Factor Models & Statistical Arbitrage
Research-led quantitative investment techniques
Finance is becoming a quantitative science.
Opening hook. Ask the audience: How does a bank decide whether to approve a loan? How does a fund decide where to invest? How is an option priced? How can a machine trade?
These look like different questions, but all use mathematics, statistics, economic reasoning, and computation. Introduce the talk as a guided journey through the major areas of quantitative finance before addressing digital innovation.
Transition: “First, what do we actually mean by quantitative finance?”
Lecturer · Quant Researcher · Quant Trader
Personal introduction (about 45 seconds). “Before we begin, allow me to briefly introduce myself. My background is in mathematics and actuarial science, and I work across teaching, quantitative research, and trading.”
“My studies have taken me through Thailand, the United Kingdom, and the United States. I also hold the Certificate in Quantitative Finance, or CQF. I serve as Vice President of the Thai Association of Quantitative Analyst and Financial Engineer.”
Transition: “This mix of mathematical theory and financial applications shapes how I approach today's discussion.”
A snapshot of quantitative finance topics from selected seminars and professional forums.
Research-led quantitative investment techniques
Methods, signals, and real-market applications
Connecting pricing methods with execution
Learning, exchange, and industry engagement
Professional context (about 30 seconds). “Beyond academic work, I regularly engage with the quantitative finance community around factor models, statistical arbitrage, options, and financial risk. The event material here illustrates those conversations and the dialogue between researchers and practitioners.”
“The common theme is turning mathematical ideas into methods we can actually test, validate, and use.”
Transition: “Let me start with a basic question: What exactly is quantitative finance?”
The application of mathematics, probability, statistics, and computation to value assets, understand markets, manage risk, and make financial decisions under uncertainty.
Models & optimization
Inference & uncertainty
Algorithms & simulation
Markets & incentives
Quantitative finance is much broader than predicting stock prices and much broader than high-frequency trading. It includes valuations, portfolio choice, risk management, credit, insurance, financial markets, and the computation behind all of these.
Explain each foundation: mathematics specifies a model; statistics estimates and tests it; computing solves or implements it; finance gives it an economic purpose.
Transition: “The many specializations can be organized around a few basic financial questions.”
Asset pricing & valuation
Financial risk management
Portfolio optimization
Quantitative trading
Credit & banking analytics
Different problems.
One quantitative language.
Read the five questions slowly, without going into mathematical details. Show that each one maps to a subfield and that a single financial institution often must answer all five.
As examples: valuation of an option; losses from a market crash; investment diversification; timing and execution; whether to approve a loan.
Good decisions require more than a single prediction.
Show how the widening fan represents uncertainty about future asset values—not a particular probability forecast. Two investments can have the same expected return but dramatically different loss exposure.
Quantitative finance studies distributions and dependencies, not only averages.
Transition: “To manage uncertainty, we need an interdisciplinary mathematical toolkit.”
Random asset prices and rates
Estimate and test models
Dynamic financial systems
Allocate limited capital
Solve difficult models
Patterns and prediction
Use simple verbal examples rather than introducing heavy equations here. Stochastic processes model evolving prices; statistics estimates relationships; differential equations describe continuous dynamics; optimization allocates capital; simulation handles intractable contracts; machine learning identifies patterns.
A single financial product may use several of these tools in sequence. Emphasize that model complexity should be driven by the financial question, not by fashion.
About 65 seconds. This map reinterprets the mathematics topics in the supplied QuantCorner material. Show the progression: calculus, linear algebra, probability and statistics form the base; stochastic processes, PDEs, time series and numerical analysis support dynamic financial models; optimization, simulation and machine learning turn models into decisions. Supplement the mathematical toolkit with financial markets, economics, risk management, and financial instruments.
The goal is not to imply that every quant must master all twelve subjects. A pricing quant may emphasize stochastic calculus and PDEs, while a statistical-arbitrage researcher may emphasize time series, inference and optimization. Financial problem first, methods second.
Transition: “The mathematics gives us a model. Programming lets us use it at scale.”
Reliable, timestamped market and institutional data.
Estimate, simulate, optimize and backtest financial models.
Run controlled systems with tests, versioning and monitoring.
Research, pricing, data science
Speed, simulation, execution
Statistics, econometrics, insurance
Market and client data queries
About 55 seconds. Based on the programming references in QuantCorner: Python, R, C++, databases, APIs and data engineering. Correct the ambiguous “Makeup Language” reference: markup or typesetting technologies such as HTML and LaTeX are useful for reporting, but they are not substitutes for programming skills.
Explain the flow from data ingestion and cleaning to model research, then safe implementation and monitoring. A beautiful equation that cannot be tested, deployed, reproduced or audited is not yet a reliable financial system.
Transition: “With these shared tools, quant professionals specialize according to the problem they solve.”
Asset pricing · Derivatives · Fixed income
Portfolio management · Quant trading
Market risk · Credit risk · Insurance · ALM
Market microstructure · Execution · Liquidity
Financial econometrics · Numerical finance · Optimization · Machine learning · Systemic risk
This is the master map for the rest of the talk. Valuation and engineering ask what instruments are worth or how contracts should be designed; investment and trading allocate capital and execute positions; institutional risk focuses on resilience; financial markets look at how orders, prices, and liquidity interact; econometrics and computational tools cut across all areas.
Take approximately two minutes to show how the categories relate. Tell the audience you will now give an example of each one.
About 55 seconds. Use four practical groupings: pricing and structuring, investment and allocation, risk and capital, and electronic trading or market making. QuantCorner also names asset pricing, portfolio optimization, quantitative investment, credit risk, algorithmic trading, market microstructure, term structure models and financial engineering. These appear here within broader functional families.
Roles and organizational titles are not standardized across employers; overlaps are normal. For example, the same researcher may build risk models, optimize a portfolio and develop machine-learning signals.
Transition: “There is another way to specialize—not by job, but by the asset being modeled.”
Stocks, indices, equity options
Bonds, interest-rate swaps
Spot FX, forwards, FX options
Energy, metals, agriculture
Corporate bonds, CDS
Tokens, crypto options, AMMs
Emission allowances, offsets
Cat bonds, weather contracts
About 65 seconds. QuantCorner specifically highlights rates and FX, equities, commodities, DeFi, carbon markets and insurance-linked products. Expand the map to eight groups by splitting rates and FX and adding credit. The labels group the most common traditional markets with emerging and specialty markets; this is a useful teaching taxonomy, not a claim that every firm uses the same desk structure.
For each market name one typical model: equities and implied volatility; rates and yield curves; FX forwards; commodity futures; credit default modeling; crypto market microstructure; carbon allowance pricing; and catastrophe bonds or weather-linked contracts.
Make the distinction explicit: “macro” refers to an approach or scale of trading, not an asset class. Likewise market making is a trading function, not an asset class.
About 75 seconds. This is the synthesis of the previous two maps. A quant is often described by both function and market: e.g., equity-derivatives pricing quant, rates trading quant, FX portfolio quant, credit-risk quant, or commodity execution quant. Walk across one row (equities) to show that four different specialist problems arise even when the underlying asset is the same. Walk down the pricing column to show that almost every market needs models for fair value.
Explain the additional “micro vs macro” distinction raised in the QuantCorner source: macro is a cross-market investment or risk perspective; microstructure studies prices and liquidity at trade/order level. A market maker may quote options, bonds, currencies or crypto.
The examples are illustrative rather than an exhaustive industry-standard taxonomy. Next, move to the technical subfields in more detail.
About 75–90 seconds. Use this slide to connect quantitative finance as an academic field to actual career opportunities. Emphasize that a “quant” is not only a trader. There are research, development, engineering, analytics, and risk-management roles.
Briefly walk down the six roles. A Quant Risk Manager measures exposures and model risk. A Quant Developer turns models into robust code and systems. A Financial Engineer designs structured products and hedging solutions. A Quant Researcher builds signals and models from data. A Quant Trader converts models into trading decisions and execution. A Quant Analyst supports valuation, analytics, and business decisions.
Close by noting why students are attracted to these careers: the work is mathematically rich, computationally modern, relevant to real financial problems, and often highly paid in competitive global markets. Then transition from careers back to the core technical subfields.
The chart highlights the expansion of quant hedge fund AUM over time, showing that systematic strategies have become a major part of the industry.
In the comparison shown, systematic equity L/S strategies outperformed fundamental peers in September 2026.
About 75–90 seconds. Use this slide as a motivational bridge between “quant careers” and the technical subfields. The key message is that quantitative finance is not only academically interesting — it is also increasingly important in real industry practice.
First, use the left chart as evidence of industry momentum. It suggests that systematic strategies represent a meaningful and growing share of hedge fund assets under management. It also shows a recent month in which systematic equity long/short returns were stronger than fundamental peers. Make clear that these are illustrative industry indicators rather than a full performance evaluation.
Then explain the structural reasons for popularity: richer data, stronger computing power, AI and automation, and the need for consistent decision systems in banks, hedge funds, asset management, fintech, and insurance. Close by noting that this popularity creates both career opportunities and deeper demand for rigorous quantitative methods.
Value is not identical to the most recent market price.
Financial assets represent claims to future cash flows or benefits. Their value depends on expected cash flows, time, discount rates, and risk.
Show the deterministic present-value formula as a starting illustration, not as a universal valuation rule. In uncertain settings we need risk adjustment and more sophisticated methods, including no-arbitrage and state-price approaches.
Examples: a government bond, a business, or a new infrastructure project.
A derivative is a contract whose payoff depends on another asset or reference variable. The pictured European call pays max(S_T-K,0) at maturity; the drawing excludes the initial premium.
Aviation companies hedge fuel exposure, exporters hedge FX movements, and investors use options to transfer risks. Pricing may require no-arbitrage arguments, replication, trees, PDEs or simulation.
Core lesson: derivatives usually transfer or reshape risk rather than making it vanish.
subject to return and portfolio constraints
Portfolio optimization chooses investment weights subject to desired returns, risk appetite, and real-world constraints such as liquidity, mandate limits, and turnover.
The schematic efficient frontier represents portfolios with the best expected return for a given risk, or the least risk for a return target, under the model assumptions.
Stress the importance of covariance and the uncertainty in estimated inputs. The mean-variance objective shown is one classical formulation, not the only one.
A loss threshold for a chosen confidence level.
Average loss beyond that tail threshold.
Losses under severe but plausible scenarios.
Explain VaR carefully: for a 95% confidence level, VaR is a 95th percentile of the loss distribution; approximately 5% of modeled outcomes have losses beyond that amount in a continuous distribution. Expected Shortfall characterizes the average loss in the tail, under the relevant definition and continuity conditions.
Stress testing asks what could happen under a specified extreme scenario instead of relying only on estimated distributional probabilities.
Emphasize model risk and non-normal tails without overloading the audience.
Systematic trading uses data-driven rules to generate trades. Walk through the pipeline: data acquisition, signal generation, risk management, execution, and post-trade evaluation.
Statistical arbitrage searches for persistent statistical relationships or systematic dislocations, but its profits are not guaranteed. Transaction fees, market impact, constraints, and regime change can eliminate an apparent historical opportunity.
A correct forecast can still lead to a poor trade at a bad execution price.
A limit order book contains resting buy and sell orders at different prices. The best bid is the highest posted buy price, and the best ask is the lowest posted sell price. Here their difference is 0.04 THB per unit.
Market microstructure studies liquidity provision, order arrivals, the bid–ask spread, transaction costs, and price discovery. This matters particularly for algorithmic trading and high-frequency strategies.
The quantities and prices are purely illustrative.
An approximate product formulation under appropriate modeling assumptions.
Use lending to demonstrate quantitative finance outside securities markets. Estimate likelihood of default, exposure at the time of default, and the fraction lost if default occurs; combine these to support expected-loss analysis.
Be technically precise: EL = PD × LGD × EAD is a standard simplified expression. More generally, expected loss is the expectation of the joint loss variables, and dependence and exposure timing can matter.
Explain how such models are used in loan pricing, lending limits, and regulatory or economic capital decisions.
Differences in repricing dates and maturities create interest-rate exposure.
Fixed income models value bonds and interest-rate products and study how rates vary by maturity. An upward-sloping and inverted curve are shown for intuition, not based on actual market data.
Asset–liability management addresses mismatches: a bank might hold long-term fixed-rate loans funded partly by shorter-term or floating-rate deposits. Rate changes can affect market value and net interest income.
Introduce modified duration as a local first-order sensitivity approximation; large shifts or complex cash flows require more advanced analysis.
Mortality and morbidity risks
Weather, yield and production losses
Floods, storms and extreme events
Actuarial finance studies the pricing and management of insurance-related risks—life, health, agriculture, and catastrophe exposures among them. Models must distinguish loss occurrence, severity, dependence, and capital needs.
Explain pooling, premium setting, reinsurance and extreme event modeling. Parametric insurance can pay according to an objective pre-agreed index, such as rainfall measured by specified stations, rather than relying solely on individual loss assessment. Such designs introduce basis risk.
This subfield closely connects quantitative finance to social and climate resilience.
Financial econometrics uses observed data to estimate financial relationships and test claims. Computational finance develops the methods to evaluate, simulate, and optimize financial decisions when analytical solutions are unavailable or impractical.
Walk through the five steps. Stress leakage-free data preparation, out-of-sample testing, parameter uncertainty, robustness and monitoring. Prediction alone is insufficient; the ultimate question is whether a model improves a real financial decision.
Designing an agricultural loan product for customers exposed to weather and income uncertainty.
Estimate borrower default risk
Determine sustainable loan terms
Control the lending portfolio
Transfer selected climate losses
Stress-test funding and losses
Estimate, validate and monitor
Use the hypothetical agricultural lending product to pull the subfields together. First credit risk models predict default likelihood; valuation and pricing calculate sustainable financing terms; optimization considers diversification and capital allocation; insurance can transfer selected climate losses; ALM and stress testing examine institutional resilience; econometrics and computation estimate, simulate, and validate every component.
None of these individual models is sufficient by itself. The practical challenge is integration and governance. Ask: “Can one product require almost all of quantitative finance?” — Yes.
Transition to digital innovation: new data, AI models and digital infrastructure can connect these processes, improve speed, and create new governance challenges.
Conclude Part I: Quantitative finance supplies foundations for asset valuation, risk measurement, capital allocation, insurance, and the operation of financial institutions. At the same time, these methods are approximations, not crystal balls. Their outputs depend on assumptions, data, structural stability, and sound governance.
Suggested closing text: “We have now seen that quantitative finance is far more than stock trading. It is the mathematical foundation of modern financial decision-making. But the data and computational technologies at our disposal are changing rapidly. The next question is not simply what quantitative finance can do, but how digital innovation will transform what it can do.”
Pause on the final question before moving into your next keynote segment.
Models of price, return, uncertainty, and constraints.
AI, APIs, cloud, programmable platforms, and streams.
Faster pricing, allocation, lending, execution, and monitoring.
Target time: 0:45. Tell the audience this is the second chapter of the keynote. The first part covered the traditional problems of quantitative finance: valuation, portfolio choice, risk, and trading. Those questions remain fundamental. The new issue is the delivery mechanism: data streams, cloud computing, AI, programmable payments, and digital distribution. Use the three boxes to explain that quantitative methods are the engine, technologies are the infrastructure, and financial decisions are the output. Transition by asking: what actually changes inside a quantitative workflow?
Periodic data → analysis → committee decision → reporting
Streaming data → model → controlled action → continuous feedback
Target time: 1:15. Introduce the digital operating loop. For example, in systematic asset management, prices and liquidity arrive continuously. Statistical models estimate risk and expected returns. Optimization translates estimates into trades. Execution systems submit orders. Monitoring detects performance deterioration or model drift. Contrast this with a once-a-month investment committee process. Stress that more frequent updates are not automatically better: bad inputs or unstable estimates propagate more quickly. Independent limits, compliance, and human escalation remain essential throughout the loop.
Predict a return distribution, volatility, regime, or factor exposure. A higher forecast score is not the same as an investable strategy.
Target time: 1:30. Explain the difference between predictive accuracy and an investment decision. An ML model may forecast short-term returns well but cause excessive turnover, concentrated holdings, or expensive trades. The risk-aware portfolio problem integrates forecasts, covariance, and trading costs. Explain each piece of the equation verbally, not symbol by symbol. Note that robust validation requires time-respecting cross-validation, realistic implementation costs, and tests during changing market regimes. The point is not AI versus quantitative finance: AI may estimate inputs, while financial economics and optimization determine what should actually be held.
Target time: 1:30. Show the bid side and the ask side. In this example, the spread is two cents. The price seen on a chart is not necessarily the price available for a large transaction. Quantitative market makers model inventory, arrival of orders, adverse selection, and latency. Traders model transaction costs and liquidity. This is one reason quantitative finance increasingly interacts with computer systems, event-driven software, and real-time data infrastructure. Remind the audience that speed by itself is not a source of risk-free profits.
Split an order across time.
Align participation with market volume.
Respond to liquidity across venues.
Target time: 1:30. Briefly explain the trade-off between trading too quickly and too slowly. Executing very fast may move the market against you. Executing very slowly exposes you to adverse price changes before the full order is completed. Institutional algorithmic execution methods such as TWAP, VWAP and liquidity-aware routing aim to balance these costs. The schematic curves are conceptual, not empirical estimates. Connect the mathematics to a practical question: could a model make money on paper but lose money after realistic execution? Absolutely, and this is why a quant trader and a quant developer must work together.
Horizon, income, liquidity
Capacity & preferences
Constrained allocation
Rebalance, taxes, costs
Drift and suitability
Target time: 1:15. Use the five-step client journey to demonstrate how classic portfolio optimization becomes a digital product. The client specifies goals, horizon, liquidity requirements and risk tolerance. Quant models propose portfolio allocations under constraints, while software automates rebalancing, ongoing monitoring and reporting. The example pie chart is purely illustrative. A well-designed system still needs suitability checks, clear fee disclosures, and human support when a client has complex needs. Digital delivery changes the reach and cost of the service; it does not eliminate investment risk or uncertainty.
Calibrate market and model inputs.
Translate Greeks into risk-adjusting trades.
Update limits, sensitivities, liquidity, and model drift.
Target time: 1:30. Explain the derivatives dashboard at a conceptual level. Delta measures first-order exposure to the underlying asset, gamma measures convexity, and vega measures sensitivity to implied volatility. A trading desk can update these measures as prices and volatilities change. The approximation on the slide illustrates how these sensitivities estimate small changes in derivative value. Real desks must manage jumps, discrete hedging, liquidity, calibration uncertainty, and trading costs. Digital infrastructure provides more timely risk visibility, while quantitative finance provides the underlying valuation and hedge models.
Transactions, income, bureau
PD, uncertainty, drift
Limits, affordability, price
Repayment and resilience
Target time: 1:45. Take a lending example from online banking or agricultural financial services. A digital platform collects permissioned information, estimates probability of default and uncertainty, and feeds these results into lending decisions. But credit assessment is not just classification. Expected loss depends on probability of default, loss given default, and exposure. A complete lending decision also reflects affordability, funding, liquidity, capital, regulation, and portfolio concentration. Explain the governance tension: alternative data can help but can also introduce proxy discrimination, privacy violations, or unstable decisions. Responsible automation includes explainability and routes for human reconsideration.
Match instructions
Compare books
Transfer assets
Linked execution can reduce reconciliation and selected counterparty or settlement frictions.
Claims, discount rates, options and default risks do not disappear.
Ownership records and contingent settlement actions become programmable.
Custody, security, interoperability, and legal finality matter.
Target time: 1:30. Introduce tokenisation as a representation of a financial claim on a programmable platform. The core insight is not speculation about token prices: it is that asset transfers and payments may be linked through delivery versus payment. The BIS 2025 and 2026 discussions explain how programmable platforms can reduce reconciliation frictions while preserving trust and final settlement. Emphasize the distinction between tokenized real-world assets, stablecoins, and digital currencies. Tokenisation does not remove duration risk, credit risk, valuation uncertainty or liquidity shortages. The economics remain; the transactional infrastructure changes. Sources: https://www.bis.org/publications/aer-2025/next-generation-monetary-financial-system and https://www.bis.org/media-releases/20260623-path-next-generation-monetary-and-financial-system-lies-safeguarding-trust-money-bis
Target time: 1:30. Contrast a central limit order book with a decentralized exchange. An automated market maker can use a formula such as x times y equals a constant, with x and y representing the token reserves. As a trade removes one asset and deposits another, the exchange rate changes. Explain price impact, liquidity provision, arbitrage, and impermanent loss in simple words. The formula is characteristic of classic constant-product pools; contemporary designs may use concentrated liquidity or more elaborate rules. Protocols remain subject to coding vulnerabilities, oracle issues, arbitrage extraction and market manipulation. Sources: https://developers.uniswap.org/docs/protocols/v2/concepts/pricing and https://www.bis.org/publications/defi-risks-and-decentralisation-illusion
Customer & merchant
Fraud / AML models
Controls & friction
Finality & liquidity
The lowest fraud rate alone is not the whole objective.
Target time: 1:15. Students may think payments are purely an IT topic, but there are important quantitative trade-offs. A fraud model needs to distinguish suspicious transactions from legitimate ones. The cost of false positives includes declined good customers and lost trust, while false negatives generate financial losses. Beyond fraud, real-time payment systems require liquidity monitoring, settlement arrangements, and operational resilience. The quantitative perspective asks us to optimize jointly over detection, customer experience, cost, and systemic safety. Explain that instant authorization does not automatically mean the same thing as final settlement.
Independent testing
Limits & overrides
Performance & drift
Accountability & audit
Target time: 1:30. For a quantitative finance audience, the digital age expands model risk beyond traditional parameter estimation. AI systems can behave poorly after regime changes. Concentration in the same cloud providers or shared AI models can generate correlated vulnerabilities. Cyber incidents and faulty real-time data can propagate into immediate financial actions. The Financial Stability Board has emphasized third-party dependency, market correlations, cyber risk, model risk, data quality and governance as vulnerabilities of AI adoption in finance. Close by highlighting four governance gates: independent validation, limits and overrides, live monitoring, and clear accountability. Source: https://www.fsb.org/2024/11/the-financial-stability-implications-of-artificial-intelligence/
Target time: 1:30. Use a simple agricultural or climate-risk example. A parametric policy pays if an objectively measured weather index crosses a threshold. The payout rule is transparent and can support faster claim administration, but a key actuarial and financial challenge is basis risk: the index may fail to match an individual customer’s actual loss. The same digital insurance landscape includes AI-assisted pricing, claims triage and fraud detection. Governance must address fairness, transparency and safety. The IAIS 2025 paper discusses governance, robustness, explainability and fair treatment in the use of AI for insurance. Source: https://www.iais.org/2025/07/the-iais-publishes-application-paper-on-the-supervision-of-artificial-intelligence/
Consented cashflow + weather
PD / LGD and uncertainty
Pricing & capital limits
Optional parametric cover
Repayment, triggers, fairness
Credit, pricing, portfolio limits and capital models.
Connected services, data pipelines and automated monitoring.
Delinquency, fairness, resilience and protection value.
Target time: 1:00. Bring the keynote back to a practical question relevant to Thailand. Imagine a digital agricultural finance platform that uses permissioned transaction information and weather data. Credit models quantify default risks, pricing models help set sustainable lending terms, optimization limits portfolio concentration, and optional parametric coverage addresses some climate shocks. The platform then monitors repayment, trigger outcomes, and potential unfairness. Highlight that this is an illustrative architecture, not an existing product or a claimed empirical result. The research agenda is to demonstrate measurable improvements in lending outcomes, customer protection, and institutional resilience.
Tomorrow’s quant is not only a model builder—but a builder of trusted financial decision systems.
Sound assumptions, validation, and uncertainty.
Secure software, scalable data, reliable execution.
Fairness, transparency, model governance and trust.
Target time: 0:45. Close Part II by summarizing that digital technology affects the data, speed, and scale of financial decisions, but mathematical rigor and accountability remain essential. Instead of ending the whole keynote here, connect this to the question for Part III: What happens when quantitative models themselves can learn from data and optimize their own actions? Introduce the next 20-minute chapter, AI in Quantitative Finance.
Returns · volatility · credit risk
Portfolio weights · trades · hedges
Out-of-sample value · costs · governance
Target: 45 seconds. This is Part III. In the first section we introduced the core mathematical questions of quantitative finance. In the second we saw how digital platforms bring financial models into live systems. Now we ask how artificial intelligence changes the models themselves. Separate three ideas: prediction of outcomes, learning actions or policies, and scientific validation. AI can change how we approximate functions, process information and implement decisions, but it cannot remove economic uncertainty or market frictions.
Opening line: “A good model predicts. A useful quant model makes decisions. A responsible quant can explain when those decisions fail.”
Regression · classification · boosted trees · neural networks
Typical output: expected return, volatility, default probability.
Actions · feedback · risk-sensitive rewards · constraints
Typical output: trade, hedge or execution decision.
Documents · retrieval · embeddings · scenario generation
Typical output: structured events, research support, simulations.
Target: 1 minute 15 seconds. Clarify that AI is not one method. Supervised learning estimates a conditional pattern, such as next-period return or probability of default. Reinforcement learning focuses on sequential actions, such as when and how to trade, where outcomes depend on earlier choices. Generative and language models can transform documents into structured signals or generate hypothetical market scenarios, but they are not automatically calibrated financial probability models.
Make the central point: a higher R-squared or classification accuracy is not necessarily a better strategy. A useful financial decision requires a loss function measured in money, capital, risk, or economically meaningful utility. We will now follow that path from prediction to implementation.
Classical linear benchmark vs. flexible ML pattern. Diagram is conceptual, not empirical performance.
Gu, Kelly & Xiu (2020), Review of Financial Studies.
Target: 1 minute 30 seconds. The canonical example is cross-sectional return prediction. For each security, collect characteristics observable at date t—momentum, valuation, size, liquidity and macro indicators. Train a machine-learning model to estimate next-period expected excess return. Relative to a linear factor regression, boosting and neural networks can capture interactions, but are also prone to overfitting. Stress that we must restrict inputs to the information available when the trade would actually be made.
A landmark empirical asset pricing study by Gu, Kelly and Xiu in the Review of Financial Studies (2020) compared methods and found that tree and neural-network models can capture nonlinear interactions relevant for estimating risk premia. It is not a claim that every neural network beats every benchmark. The output is a return estimate, not an executable strategy. Portfolio construction and trading costs come next. Source: https://doi.org/10.1093/rfs/hhaa009
Learning nonlinear time dependence is useful only if regime changes can be detected and managed out of sample.
Interpret probabilities as model outputs, not knowledge of future regimes.
Target: 1 minute 20 seconds. Financial data are not stationary. Volatility can jump, correlations can rise sharply, and past relationships can collapse. Time-series ML seeks richer dynamics for volatility forecasting, detection of latent market regimes, and dependence modeling. The figure is an illustrative normal period followed by a stressed period; it is not historical data.
Give the audience two contrasts. Traditional GARCH models describe conditional volatility through an interpretable structure. More flexible models can use option-implied information, realized volatility, liquidity and macroeconomic signals together. But regime probabilities depend heavily on the information set and assumed model. Never promise that the model knows when a crisis begins. The economic use is adaptive risk limits, position sizing and scenario weights.
Target: 1 minute 15 seconds. Finance has enormous unstructured information: disclosures, earnings calls, rating reports, policy releases and news. Language models can turn documents into structured events, sentiment or embeddings, which are then inputs to quant strategies or risk-monitoring systems. They may also assist with research summaries; do not equate plausible summaries with verified facts.
Emphasize information timestamps. Suppose a company files an earnings release at 08:30. A backtest must not use the extracted signal before the document became public and sufficiently available for processing. Revised transcripts, modern text databases with hindsight cleaning, and company identity mistakes are common leak pathways. Test the incremental economic value against simpler lexical, sentiment and macro benchmarks. Human validation and provenance remain crucial.
Penalize volatility and costly rebalancing.
Budget · exposure · liquidity · leverage.
Target: 1 minute 30 seconds. Explain why forecasting is only one stage. Suppose the machine-learning model predicts expected returns for 300 stocks. Holding every stock in proportion to the forecast would ignore risk, diversification, costs and legal or policy limits. The portfolio optimizer trades off expected return against covariance risk and transaction costs. The displayed equation is a stylized quadratic mean–variance objective with an L1 turnover penalty.
The sum of weights and additional constraints must still be specified, for example budget = 1, leverage limits, sector caps, liquidity restrictions and transaction limits. Robust portfolio construction also responds to prediction uncertainty; small errors in expected returns can cause unstable allocations. The practical question becomes: does AI produce a better risk-adjusted net portfolio than a transparent, cost-aware baseline?
Lower fees / adverse selection risk.
Faster fill / potential impact.
Avoid immediate cost / timing risk.
Lower inventory and footprint.
Test in simulators and historical replay. Compare with TWAP, VWAP and implementation-shortfall baselines; simulation gaps matter.
Target: 1 minute 30 seconds. Explain the difference between a price forecast and a trading policy. The agent observes the current market state: bid–ask spread, order-book depth, remaining quantity and time to completion. It chooses whether to post a limit order, take liquidity with a market order, reduce size or wait. The reward captures economic benefit after transaction costs and a penalty for risk or incomplete execution.
Reinforcement learning is attractive because execution is a sequential stochastic-control problem. But an agent may overfit its simulated market, exploit bugs, or learn behavior that does not survive live liquidity conditions. Restrict actions and market impact, test conservative baselines including TWAP and VWAP, monitor safety constraints, and require governance approval before any live deployment. Explain that an accurate forecast is neither necessary nor sufficient for good execution.
Target: 1 minute 30 seconds. Compare classical derivative hedging with learning a hedging policy directly. The standard delta rule is based on the sensitivity of derivative value to the underlying and a model for the option. It can be an excellent baseline in liquid markets and short rebalancing intervals, but real hedging faces discrete trades, transaction costs, liquidity constraints, risk limits and model uncertainty.
Deep hedging formulates a stochastic control problem: choose dynamic hedging actions to minimize a risk functional of hedging loss plus trading costs. The formula on the slide separates a risk measure rho applied to hedging loss from expected trading costs; do not imply all risk measures are expectations. Cite Bühler, Gonon, Teichmann and Wood, “Deep hedging,” Quantitative Finance 19(8), 2019, DOI 10.1080/14697688.2019.1571683. Note that learning does not remove model risk: it changes where assumptions enter and how we validate them. Source: https://www.tandfonline.com/doi/full/10.1080/14697688.2019.1571683
Visual is synthetic and uncalibrated: it does not establish that a model learns extreme losses.
Target: 1 minute 15 seconds. Generative models such as GANs, diffusion models and autoencoders can be used to sample synthetic financial returns or market states. They are useful for augmenting scenarios or studying how strategies react to unusual joint price moves. But generated samples must not be misrepresented as real financial history, and a model trained on typical days may miss crashes or liquidity spirals.
Show the scenario cloud as entirely illustrative and uncalibrated. Ask whether it reproduces fat tails, tail dependence, volatility clustering, conditional relationships and realistic execution costs. One useful approach is to combine financial structural constraints with flexible generators, then benchmark their scenarios against empirical and stress-test targets. This is a research opportunity, not evidence that any AI generator forecasts the next crisis.
ES identity shown for continuous loss distributions.
Coverage calibration, exceedance clustering, stress resilience and decision usefulness.
Target: 1 minute 15 seconds. Move beyond predicting average returns. A quant risk manager needs loss distributions, tail quantiles and expected shortfall. AI methods can estimate nonlinear conditional risk from asset positions, factors, macro signals and liquidity, but it remains crucial to test coverage: does a nominal 99% VaR get exceeded at the expected rate over meaningful evaluation periods? Are failures clustered during market stress?
Explain the slide formulas using portfolio loss L, with alpha as the confidence level. The conditional expectation identity for Expected Shortfall is stated for continuous loss distributions; otherwise its general quantile-average definition is required. An impressive-looking fit to ordinary periods cannot establish crash protection. Backtest exceedance frequencies and stress-test joint behavior, then connect risk forecasts to position limits and capital decisions.
| Asset class | Typical input | AI output | Quant decision |
|---|---|---|---|
| Equities | Factors, news, order book | Expected returns / liquidity | Stock allocation and execution |
| Fixed income | Yield curves, macro, flows | Curve and spread dynamics | Duration and hedging |
| Foreign exchange | Rates, order flow, macro | Conditional FX risk | Hedge ratio and trade timing |
| Credit | Firm fundamentals, payment data | PD and loss risk | Lending limits and pricing |
| Commodities | Weather, inventories, futures | Basis / term-structure regimes | Physical and financial hedges |
| Digital assets | Trades, funding, on-chain events | Liquidity and liquidation risk | Exposure and execution limits |
Target: 1 minute 10 seconds. Reconnect to the asset-class taxonomy from Part I. The underlying AI technique might be a boosted tree, transformer or reinforcement learner, but the economically relevant target changes with each asset class. Equity quants may use characteristics, fundamental text and market microstructure. Rates desks care about curve dynamics and hedging. FX trading links exchange rates to interest differentials and flows. Credit teams forecast defaults and loss severities. Commodity risks depend on physical inventory and weather. Digital assets introduce funding, liquidation and fragmented liquidity dynamics.
Emphasize that asset class expertise remains indispensable. An identical model architecture does not imply identical training data, payoff mechanics, constraints, frequencies or validation.
Search literature, annotate filings, draft research code.
Source validation, sandboxed tools, permission limits.
No unattended orders, confidential data leaks or money movement.
Target: 1 minute 10 seconds. Language-model agents can combine retrieval, coding and analysis tools to help a quantitative research team: summarize public disclosures, run preliminary diagnostics, document a model and compare a result with prior work. But they may hallucinate numerical data, misread sources, or follow malicious instructions embedded in retrieved content.
A safe architecture restricts access to approved sources, versions all data and scripts, validates outputs with deterministic tests, and requires explicit human authorization before important financial actions. Distinguish a research assistant from an automated trader. The model's ability to write persuasive text is not evidence of a true signal or legal trading authority. Include cybersecurity, confidential information and data minimization in the evaluation.
Fit parameters on past dates
Tune only on prior dates
Untouched forward period
Refit on newly available history
Point-in-time data and actual publication delays.
Purge / embargo where overlapping windows contaminate splits.
Fees, spread, slippage, impact and financing.
Account for many trials, false discovery and drift.
Target: 1 minute 50 seconds. This is the most important methodological slide. Machine learning can overfit years of financial history with subtle leakage: survivorship bias in stock universes, revisions to fundamentals, financial news known only later, labels that overlap training and test windows, and repeated searching for the best hyperparameters. A rolling or expanding training window evaluated on subsequent untouched observations better approximates real model deployment.
Explain the four colored stages: train, validate, test, and roll forward. Purging and embargo gaps can help when event labels or prediction windows overlap across splits. Economic validation matters: gross returns are reduced by commissions, bid–ask spreads, slippage, market impact, borrow costs and financing. In many strategies, capacity and turnover dominate headline predictability. Benchmark against sensible linear factors, no-ML strategies, and simple execution rules. When many variants have been tested, account for selection bias and false discoveries.
Finishing line: “The best financial AI is not the model that looks smartest on yesterday's data; it is the one that survives tomorrow’s data, after costs.”
Markets change; calibration and signal strength decay.
Shared models or signals may intensify crowding.
Dependence on the same cloud and model infrastructure.
Data poisoning, prompt injection, fraud and outages.
Target: 1 minute 30 seconds. Rapid AI adoption can improve efficiency and analytics, but it can create shared vulnerabilities. The Financial Stability Board has highlighted third-party and provider concentration, correlated market behavior, cybersecurity, data quality and model risk as potential financial stability concerns. IOSCO has examined AI use cases and associated investor-protection and market-integrity risks in capital markets. The 2025 FSB follow-up discusses how to monitor adoption and such vulnerabilities. Sources: https://www.fsb.org/2024/11/the-financial-stability-implications-of-artificial-intelligence/ ; https://www.fsb.org/2025/10/monitoring-adoption-of-artificial-intelligence-and-related-vulnerabilities-in-the-financial-sector/ ; https://www.iosco.org/library/pubdocs/pdf/IOSCOPD788.pdf
Highlight a governance process: define intended use, record ownership and provenance, independently validate, set trading and lending controls, monitor drift, and preserve manual stop / override authority. If mentioning US model risk regulation, reference the Federal Reserve's revised 2026 guidance, SR 26-2, which superseded SR 11-7: https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm . Do not assume this US supervisory rule applies to Thailand; the general governance principles are transferable.
Target: 1 minute 15 seconds. Conclude the AI chapter by connecting all three parts of the keynote. Quantitative finance gives us the mathematical foundations and economic objectives; digital finance provides the data and implementation infrastructure; AI extends the ways we learn relationships and optimize actions. None of these can replace validation and human accountability.
For students and young researchers, the strongest career portfolio combines probability and stochastic processes, statistical learning, programming, optimization, and a deep understanding of financial markets. For institutions, responsible innovation means investing in data governance, experimentation, scientific benchmarking, controls and independent validation. This then leads to a concrete research case: how can stablecoin lending rates control a coupled peg-and-liquidity spiral?
Transition to Part IV: "Let me show how quantitative modeling, digital financial infrastructure and risk-aware control come together in one of our stablecoin research projects."
Peg, lending liquidity and trading interact.
How strong and how fast should rates respond?
Can policy reduce the chance of a real depeg event?
Target 45 seconds. Introduce the research article by Nassamon Bootwisas and Pasin Marupanthorn. This study links three ideas from the earlier keynote: decentralized finance, dynamic quantitative risk models, and optimal control. Stablecoins aim to remain close to a reference price, but decentralized lending changes both borrower incentives and trading pressure. The paper asks when a peg-sensitive rate is stabilizing, and when feedback or delayed information makes matters worse.
The contribution is a theoretical framework supported by entirely synthetic, reproducible computations. It is not a historical case study or a deployed-protocol recommendation.
Cash available inside the lending pool to satisfy withdrawals and new borrowing.
Trading depth available for buying or selling the stablecoin in external venues.
Target 1 minute 10 seconds. Start with a small fall in the stablecoin market price. Token holders may sell, borrowers and liquidity suppliers change their actions, and available cash in the lending reserve can tighten. A lending-rate increase may reduce debt demand and help the peg, but it also affects borrower behavior, repayments, funding, and the external market. Those reactions are endogenous.
Distinguish reserve liquidity—the cash in the lending pool—from market liquidity, which is the depth available to absorb trades. A pool can have substantial accounting assets but limited available cash, while the token may trade in a shallow AMM. The paper models this loop explicitly rather than assuming a higher borrowing rate necessarily restores stability.
Events create discontinuous cash/debt and market changes; oracle observations and rate quotes update on their own schedules.
Target 1 minute 5 seconds. Here is the concrete modelling innovation. Let C be available cash and B be variable-rate debt; reserve utilization U is B divided by C plus B. A borrower taking new cash decreases C and increases B, whereas a supply deposit raises C. A token-based debt cancellation is not the same accounting operation as a cash repayment.
Outside the reserve, stablecoins trade in a constant-product market; the price is written P=exp(-x), so an increase in x indicates a deeper discount. Each accepted transaction can be associated with an external buy or sell, which changes the price discontinuously. The stochastic state also tracks issuer shocks, market depth, an oracle filter, cached observations and quote clocks. This is an idealized but explicit hybrid financial system, not a claim to replicate every feature of a live protocol.
Target 1 minute 10 seconds. The paper proposes an extra, capped premium for a lending protocol. The first factor is a utilization gate: when the reserve is comfortably liquid, the premium can be zero, but it ramps up as utilization enters a stressed range. The second factor is a positive peg-discount signal above a deadband. The gain k determines sensitivity, and the cap prevents an unbounded quote.
The oracle signal may be filtered or sampled, so the quote is based on the information actually available to the lending protocol. A very important distinction is that Aave's inspected utilization-based rule does not itself implement this proposed peg premium. The paper analyzes a possible extension, under stated assumptions.
A peg discount can amplify liquidity stress.
Reserve pressure can generate selling or lost peg support.
The filter does not observe or react instantaneously.
Target 1 minute 10 seconds. For mathematical analysis we reduce the dynamics to three risk coordinates. The first is scarcity v, reflecting the reserve's debt-to-cash pressure. The second is the log price discount x. The third is the filtered observable signal S. The rate premium r feeds into both reserve pressure and market price. The parameter eta is net liquidity relief, including interest accrual; chi is the peg response and could be negative. The feedback coefficients b and c make the two financial states mutually dependent.
Additive Brownian and compound-Poisson shocks represent uncertainty in the analytical closure. These coefficients are theoretical assumptions and the numerical values later are synthetic. This closure enables exact local stability calculations; the paper returns to the accounting-level hybrid process for first-passage and policy results.
Approximately 0.231 < k < 6.510.
Needed here for the active, unsaturated equilibrium with the illustrative 200-bp cap. Gate consistency is also required.
Target 1 minute 20 seconds. This is a main analytical finding. For a specified three-state stress model, the linearized dynamics are locally stable only inside a finite interval of the peg-feedback gain. In the paper's numerical example, the exact gain bounds are three thirteenths and three hundred nineteen forty-ninths. Below the lower threshold, reinforcing feedback dominates. Above the upper threshold, filtering can create an oscillatory linear instability.
The stability calculation comes from the cubic characteristic polynomial and its Hurwitz conditions. Importantly, matrix stability alone is not enough. For a two-percentage-point annual premium cap and the chosen model loads, active unsaturated equilibrium also requires k greater than six elevenths. The validity of the utilization gate and the cap must be checked separately. None of these numbers is an actual Aave parameter recommendation.
Some states require liquidity or balance-sheet intervention; merely raising the borrowing rate is insufficient.
Assumes b c > a d, nonnegative loads, χ ≥ 0 and the positive unstable eigenmode of the stated closure.
Target 1 minute 5 seconds. A subtle but important result is that even a valid rate cap cannot guarantee global stabilization. Suppose scarcity and peg discount reinforce each other more strongly than their natural mean reversion. The model then has a positive real eigenmode. If initial stress exceeds a threshold determined by the cap and the eigenvectors, all allowed bounded-rate policies fail to arrest a deterministic liquidity spiral.
This is not a claim that all stablecoins inevitably collapse. It is a conditional result for a supercritical, specified closure with nonnegative loads and the stated sign of the policy response. Financially, it says that an interest-rate tool has limits. Beyond those limits, liquidity provisioning, risk reduction, or other balance-sheet interventions may be necessary.
Continuously observed peg input
Current observation held between updates
One-epoch stale observation
Target 1 minute 10 seconds. The previous slide concerned the strength of the response; this slide concerns the timing of the information. Continuous observation, a held observation, and a one-period stale observation produce different exact transition matrices at decision epochs. For the chosen synthetic example with gain k equal to three, the first computed loss of discrete sampled stability appears near 13.969 hours for continuous information, 6.631 hours for a held current observation, and only 3.372 hours for one-period stale input.
Use the chart to show how the spectral radius crosses one sooner as the oracle becomes less timely. These are example-specific first crossings, not universal clock limits or a claim that every longer interval must be unstable. A credible financial controller must be validated using the actual latency and update policy.
FJ = FP ∩ FU. The union means either event and is a different risk metric.
Target 1 minute 5 seconds. Risk management should care whether a stablecoin actually crosses an unacceptable price or utilization threshold, not merely about the expected price next week. We distinguish three events. Depeg occurs when the log-discount crosses a threshold. Liquidity stress occurs when utilization becomes too high. Joint stress requires both at the same time. The union, in contrast, counts either event.
The paper uses a killed Markov process: once a path enters the selected failure set, the remaining probability mass is killed. Its missing mass is the cumulative first-passage probability. This construction counts exact transaction jumps that overshoot the boundary and, in the stipulated Brownian between-event model, bridge crossings between numerical observation points. The specified failure set matters: reducing joint stress need not reduce standalone depeg probability.
Probability of crossing a failure threshold before the next decision.
Nonnegative cost of charging an additional premium.
Probability-weighted future risk on surviving paths.
The Bellman claim is for the specified transition law, candidate actions and decision schedule.
A nonnegative baseline-risk residual yields a weak guarantee; strict improvement needs positive visited residual.
Target 1 minute 15 seconds. Our fourth theorem gives a genuine finite-horizon decision criterion. At each decision time, the controller chooses among finitely many permitted premiums. The Bellman recursion adds three components: the probability of failure before the next decision, the nonnegative premium burden, and the continuation value on paths that survive.
The zero-premium baseline action is always feasible and has zero additional burden. Therefore the optimal risk-plus-burden value cannot exceed the baseline failure probability. Since burden is nonnegative, optimal pure failure probability is also no higher than baseline within this model. For a proposed preset policy, a residual comparing its one-step performance with baseline future risk provides a separate certificate. That residual must be nonnegative over all live states; a few favourable grid nodes do not certify the continuum. Different failure metrics require different killed kernels.
Risk-aware policy burden: 0.03033; total modeled objective: 0.16932. Burden is not a dollar cost without exposure conversion.
Target 1 minute 20 seconds. Here are actual numerical outputs from the manuscript, but they are from a synthetic finite-state scalar approximation, not live DeFi data. Over seven days the base policy has estimated depeg risk 24.59 percent. A fixed preset gate has 13.24 percent, while the risk-aware Bellman quote has 13.90 percent. Why is the optimal policy not the one with the lowest failure probability? Because the objective contains both the chance of failure and a policy penalty. The optimized controller accepts slightly higher risk in exchange for lower rate burden.
The right panel shows the premium is dynamic: it depends on observed log discount and how much time remains to maturity. The risk-aware policy has burden 0.03033 and total objective 0.16932 in the finite-state numerical model. The exact Bellman theorem applies to a specified kernel, but these plotted risk numbers are numerical approximations and do not certify continuous-state discretization error.
Target 1 minute 35 seconds. This is the most economically important numerical comparison. The full accounting-level hybrid simulator includes cash and debt resets, external trades, issuer regimes, and lagged oracle observations. In each response scenario, the study uses 4,000 paired paths for the baseline and gate arms.
In the aligned-response scenario, the joint probability of a one-percent depeg and utilization above 90 percent falls from 22.40 to 21.12 percent. The paired difference is minus 1.275 percentage points, with a 95 percent interval from minus 1.649 to minus 0.901. In a no-response scenario the policy changes nothing: both are 25.57 percent. With an adverse supply response, the point estimate rises from 24.02 to 24.12 percent; the interval for that difference crosses zero, so it is not evidence of a definite worsening.
Notice the standalone depeg probability stays around 42.2 percent in both arms. Improvement in joint stress is not the same as improvement in stand-alone peg risk. A preset rate gate can be effective only if its economic transmission matches the model's assumptions. All values are synthetic and should not be presented as historical backtests or live Aave results.
Check active stability, the utilization gate and cap consistency separately.
Held and stale information can change the system's stability boundary.
Measure depeg, liquidity and joint-stress risk separately, including jump overshoots.
Verify market and transaction responses, model law errors and numerical approximation.
Target 50 seconds. Close with four lessons. First, a stronger policy is not necessarily better: evaluate both a lower and an upper gain threshold and ensure the controller remains within its permitted branch and cap. Second, use the real observation schedule, because stale oracle inputs may invalidate a seemingly safe continuous-time model. Third, judge policy by actual failure events, not only by a favorable drift or utilization response. Fourth, move from synthetic modeling to empirical calibration and independent implementation testing before deployment.
The package documents conventional proof auditing and reproducible computational checks; its Lean formalization candidates have not been executed or machine verified. The full high-dimensional accounting-level Bellman controller has not been numerically solved. This makes the paper a framework for conditional risk control, not evidence of universal stablecoin stabilization. End with the message that sound quantitative finance couples rigorous models with realistic financial mechanisms and honest uncertainty.
Transition to Part V: Our next study complements this stability model with empirical data on how public DeFi insurance markets select protocols and allocate risk-transfer capacity.
When a protocol becomes publicly coverable
Lagged economic activity before listing
Available cover capital at snapshot
Quoted cover prices, not safety ratings
Target 45 seconds. This section presents the empirical paper “Digital Insurance and Protocol Insurability in Decentralized Finance: Longitudinal Evidence from Public Cover Listings.” Unlike our previous stablecoin research, which developed a theoretical risk-control mechanism and synthetic experiments, this study examines observed public insurance listings and protocol histories.
The central question is what a public cover listing really signals. Is it evidence of technical safety? Or does it reflect a protocol that has become economically significant, visible to potential buyers, and assessable by the cover market? We will use public listing events, lagged total value locked, incident histories, available cover capacity, and annual quoted prices to answer that question. Source: supplied manuscript, abstract and pp. 4–8.
Target 1 minute 5 seconds. Explain the paper’s institutional distinction using Nexus Mutual. Product listing is the first layer: the protocol has become a named risk object in a public cover market. Capacity is a different layer: the amount of cover that can currently be supplied. An active listing can coexist with little or zero capacity. Price is yet another layer: the quoted annual cost, influenced by target rates, staking allocation, utilization, purchases, and the product’s rules.
These three public observables give us unusual visibility into underwriting decisions that are mostly private in traditional insurance. They are not purchased cover, premiums actually paid, claim volumes, or a security certification. Source: manuscript §2.1 and §3.3, pp. 5–8.
Target 1 minute. The four hypotheses map an insurance question into observable digital-finance data. First, protocol scale: a large protocol may attract potential customers and make the financial risk worth assessing. Second, age: long operating history could help risk evaluation, but the insurer may prioritize recently launched protocols with strong interest. Third, incidents: an exploit is bad risk information but may also raise users’ demand for cover. Fourth, capacity: once a protocol is listed, the willingness and ability to offer capital can be studied separately from the act of listing.
We do not assume these characteristics mechanically measure code security. Source: manuscript §2.2, pp. 6–7.
Target 1 minute 15 seconds. This is the reproducibility backbone. The 21 July 2026 Nexus snapshot contains 456 products, of which 170 are public, active, and non-private. A narrower set of 94 products is classified as direct protocol cover, and only 74 have sufficiently precise reviewed links to DeFiLlama protocol identities. This is important because depeg, custody, vault, and portfolio cover are not the same insured object as direct protocol cover.
The historical study uses Ethereum ProductSet events, identifies 31 direct-protocol first listings between September 2024 and June 2026, excludes four events without prior-month TVL, and applies the US$1 million activity rule. The resulting main risk set contains 25 events and 14,284 protocol-months across 854 protocols. Forty-five protocols are already covered at baseline and are left-censored. Variables enter only at the previous month-end. Source: manuscript §3 and Table 1, pp. 7–11.
Target 1 minute 10 seconds. We model the probability that a protocol receives its first listing in month t, conditional on not previously being listed. The key predictor is total value locked from the preceding month-end, avoiding contemporaneous look-ahead. The model also includes lagged age, whether a public exploit happened before that date, broad risk groups, and calendar-quarter controls. The complementary log-log link suits a discrete-time event-history formulation, and errors are clustered at protocol level.
A hazard ratio is an association in the listing risk set. Exponentiating the coefficient on log TVL gives the relative hazard for an approximately 2.72-fold increase in TVL. The authors also use a logit robustness check. This is not the probability of insurance paying out, and it does not identify causal impact. Source: manuscript §4.1, pp. 9–10.
Target 1 minute 25 seconds. Focus the audience on the primary result. In the US$1 million activity panel, the log-lagged-TVL coefficient is 0.297 with a clustered standard error of 0.115. The hazard ratio is 1.345, with 95 percent interval 1.073 to 1.687 and p equals 0.010. Thus a roughly e-fold increase in prior-month TVL is associated with a 34.5 percent higher monthly hazard of receiving the first listing, conditional on controls. The logit robustness model gives nearly the same result.
But threshold sensitivity matters. The hazard ratio is 1.286 at US$0.5 million (p .038), 1.345 at US$1 million (p .010), and 1.318 at US$5 million (p .053). With no threshold and near-dormant observations included, the result drops to 1.106 with p .551. Make explicit that a visible market-selection signal exists in an economically active subset, not across every protocol regardless of activity. Source: Table 2 and Figure 2, manuscript pp. 11–14.
Target 1 minute. The age effect is strong and negative in the listing model: the reported hazard ratio on log age is 0.133, p below .001. This does not mean older code is less safe. Conditional on having meaningful scale, the observed listing expansion disproportionately targets newer protocols. Expansion strategies, anticipated buyer demand, left-censoring, and cohort effects are possible explanations.
Prior public exploits yield a hazard ratio of 1.571, but the interval ranges from 0.172 to 14.365 and p equals .689. It is far too imprecise to claim that hacks raise or lower the listing hazard. Emphasize competing underwriting concern and demand-attention channels. Source: manuscript §5.1 and Table 2, pp. 11–12.
Target 1 minute 5 seconds. Figure 1 shows the 31 dated first-listing events, with a cluster of seven during October 2025, and the median protocol TVL before and after listing. TVL is indexed to 100 one month before listing. At month minus two it is 94.7, rises to 117.9 in the listing month and 120.3 two months afterward, then generally returns toward the prior level while dispersion broadens.
This is not evidence that obtaining cover caused protocol growth: the event study is descriptive, the support varies over time, and there is no untreated counterfactual. It may instead reflect anticipated demand, launches, or market cycles. Source: manuscript Figure 1 and §5.1, pp. 12–13.
Target 1 minute 10 seconds. The current cross-section answers who is listed now rather than who first receives a listing next month. It contains 69 insurable protocols with reviewed links and 841 supported controls. Propensity matching retains 61 pairs and balances all measured covariates to absolute standardized mean differences below .10, with maximum .061. Matching is a balance diagnostic, not a causal treatment-effect design.
The logit availability model finds positive average marginal effects for log total value locked, .024; log chain count, .037; and reported audit status, .070. These reflect current associations only. Reported audits and chain breadth were not reconstructed historically, so they are not claimed as lagged longitudinal predictors. Source: manuscript Table 3, Table 4, §5.2, pp. 14–16.
Target 1 minute 15 seconds. Among the 69 protocols with reviewed current cover, the conservative capacity measure takes the maximum linked-product capacity to reduce obvious double counting. The median insurable protocol has TVL US$61.53 million and conservative available cover capacity US$4.76 million. In the log-capacity equation, log TVL has coefficient 1.080, HC3 robust statistic 3.73, and the estimate is stable to winsorization. In the annual-price equation, log TVL has coefficient minus .539, robust statistic minus 4.74.
These are cross-sectional associations. Available capacity is not cover purchased; price is the current quote, not a pure actuarial measure of loss costs or claims. Cover assets, staking allocations, utilization, target rates, and buyer demand also influence these terms. Figure 3 makes the dispersion visible. Source: manuscript §4.3, §5.3, Table 4 and Figure 3, pp. 10, 15–16.
Target 1 minute. This is the integrity slide. The study’s strengths are reviewed event dates, exact entity linkage, lagged exposures, clustered inference, activity-threshold sensitivity, matching diagnostics, and transparent reporting of exclusions. But temporal ordering does not remove unobserved underwriting decisions, anticipated product launches, or demand salience. The primary event study has 25 events, limiting precise estimation for infrequent covariates such as hacks.
The longitudinal universe could miss protocols that disappeared before collection. Available capacity and prices are cross-sectional, with possible shared pool constraints. We do not have cover purchases, premiums paid, claim outcomes, or internal underwriting deliberations. Therefore this paper cannot certify the safety of any protocol. Source: manuscript §6, pp. 17–18.
Target 1 minute 15 seconds. Close the research-to-practice gap with a careful four-step workflow. First, identify exactly what object is being covered: a specific protocol, version, token depeg, vault, or custody relationship may each have different risks. Second, assess protocol fundamentals independently: audits, governance, oracle and bridge architecture, liquidity, leverage, incident response, and composability. Third, inspect actual cover capacity, price, exclusions, waiting periods and claim-assessment rules. Fourth, continuously monitor changing market and product conditions.
Explain that cover availability may help teams discover market attention and the assessability of an exposure, but it must never be treated as an objective protocol safety score or consumer recommendation. The workflow shown is a proposed practical interpretation, not a newly estimated causal effect. Source: manuscript §7.2 and §6.4, pp. 18–19.
Scale matters primarily among economically active protocols.
Larger covered protocols generally show more available capacity and lower prices.
Market attention, demand and risk-based selection can coexist.
Target 1 minute 35 seconds. Summarize three lessons. First, digital insurance provides a newly observable selection mechanism: in the relevant active sample, larger protocols are more likely to receive their first listing, but the association vanishes when dormant protocols are included. Second, we should separate coverage availability from underwriting depth and price. Within current coverable protocols, larger TVL is associated with more supply capacity and lower annual quotes. Third, none of these results is a direct score of code quality or a guarantee of claim settlement.
The broader contribution is a reproducible longitudinal link between smart-contract listing events, a manually reviewed protocol taxonomy, historical activity, public incident data, and current cover-market terms. Future work should cover multiple insurers and capture historical purchase flows, premiums, claims, capacity changes, deprecation, governance choices, and shared risk dependencies. End by connecting to the larger keynote: quantitative finance supplies the statistical discipline, blockchain creates transparent event records, digital insurance creates a market for risk transfer, and AI may eventually help with assessment—but validation and economic interpretation still matter. Source: manuscript §7, pp. 18–19.
← → / Space — Change slides
N — Toggle speaker notes
O — Show all slides
F — Toggle fullscreen
T — Start / pause rehearsal timer
Home End — First / last slide
Esc — Close an overlay
You can use the print button and choose Save as PDF. No internet connection or additional assets are needed.