Inside the ACM Winter School: AI, Finance, and Learning

ACM Winter School on AI and Finance: What I Learned and Why It Matters

AI and Finance are currently booming domains across the globe. Countries and institutions are trying to make citizens AI-literate through various learning programs. I am an Undergrad student set out to learn new technologies that will equip me to contribute meaningfully to society. I was interested in finance because of Time Research it felt a very cool idea to deal with money, markets and decision-making made me venture into Data Science. Later I explored a lot of stuff online to understand how my interest in solving problems can help me to contribute in the domain of Finance. However, it was very difficult. I was unable to understand the literature present, there were many new financial terms coming up and the more I tried the more I felt that I was either directionless or simply stupid. Then after 2 years with ACM I came across the Winter School in AI and Finance organized by IIIT Hyderabad and Alpha Grep. I decided to apply, hoping this would give me the structured foundation I was missing. When I received the selection email, I was genuinely elated. During a phase when I was struggling to break into finance, this opportunity felt like the stepping stone I truly needed.

Day - 1

The first day we started out on exploring how finance actually started. I came across lot of names from the history but was advised to remember few names like Ito Kiyoshi, Markowitz, Sharpe and Black Scholes. Later we tried to understand what are forward contracts and options, got an overview on option pricing.

We also went through what are the kind of financial instruments are there (Risky and Risk-free) had to remember examples like stocks and bonds which were used throughout the course. There were other key terminologies defined like arbitrage, No Arbitrage Principal, Long and Short positions etc. Then came an entirely different subject which I never knew-Stochastic Calculus this mathematical treatment had actually helped me to understand the terminologies defined earlier better and it was a lot of fun learning something challenging. Mathematics help spin off stories, Stochastics Calculus actually helps beginners understand financial terms better through it. Then we formally defined what Brownian Motion is exactly and from there we discussed properties of Brownian Motion, Quadratic Variation(QV) and First Order Variation(FOV). Then we defined Ito's Integral and understood its properties. This was mathematically heavy but we formalize the financial terms that help to further understand the course better. We were made to do a hands-on session on Neural Networks. This was a refresher for me touching upon the basics on ML and NN that will help all the students to get on same page.

Day - 2

The second day we continued with stochastic calculus. We defined Ito's process, solved the Black Scholes equation and appreciated it. We understood the call and put options through Indigo stock market prices(The Indigo crisis was presently happening so it became a highly relevant example to discuss upon). We were introduced to Portfolio Theory, Two asset theory and how it extends to Two fund theory this session became interesting because we had some prior understanding of the various terminologies which were discussed in Day-1. Then we had a hands on session on the same this lab helped to understand on how one can capture uncertainty in the financial market, Quantify risk using variance and covariance. I was able to understand the Two asset pricing concept better in the hands on, I observed that by varying asset weights and correlations combining assets reduces the overall portfolio risk and Lower correlation leads to better diversification benefits. This enabled me to imbibe that diversification is a powerful concept in finance. We then extended these ideas to multi-asset portfolios, where the learning became even more insightful. By simulating multiple combinations of assets, we observed the emergence of the Markowitz bullet, the characteristic shape formed when expected return is plotted against risk. This visualization clearly demonstrated how different portfolios occupy different positions in the risk–return space, and that only a subset of these portfolios are truly optimal. Watching the Markowitz bullet take shape through experimentation made modern portfolio theory feel concrete and intuitive rather than purely mathematical. From this set of portfolios, we identified the minimum-variance portfolio, which represents the lowest achievable risk for a given universe of assets. Building on this, we traced the efficient frontier, capturing portfolios that either maximize expected return for a given level of risk or minimize risk for a desired return. This exercise clarified how rational investors can systematically choose optimal portfolios based on well-defined criteria instead of relying on intuition or ad hoc decisions. To connect portfolio theory with real market behavior, we performed CAPM-style regressions. Through this, I learnt how to estimate beta, which measures an asset’s sensitivity to market movements, interpret beta as a measure of systematic risk, and understand what alpha and residuals reveal about asset performance beyond market effects. This analysis helped bridge theoretical models with practical market dynamics, offering insight into how individual assets behave relative to the broader market.

Day - 3

We kicked off the third day Sequential Decision Making. We were introduced to Reinforcement Learning, I had no idea what it was so it was a bit difficult to understand the basic intuition however, the professor simplified this in a very systematic manner by visualizing a problem (Eg: Finding Highest possible 4-digit number from given random numbers) as multiple single stage dynamic programming problem. We then proceeded to defining RL setup which is Markov Decision Process and the terminology we will be following for this course like policy, agent-environment interaction, finite and infinite horizon. The session then shifted to input modeling, where we explored fitting distributions to financial data and discussed why simulation is essential. We discussed Bootstrap(like Random Sampling) and then tried to understand why exactly it doesn't work and then proceeded to actually define the design thinking principles to impose a distribution which are

Adequacy of features
Simulatability

We saw how inverse transforms can help in simulating univariate features and the same can be extended to multivariate features. we had session on copulas (non-parametric model) for simulating in a data-driven manner. We further went on defining Sklar's Theorem, Gaussian Copula and Empirical Copula. We then got a brief overview about model risk.

Day - 4

We continued with RL by defining Q-value function for a given policy. Defining Value policy and understand model based algorithms to find an optimal policy by Value Iteration(VI), Policy Iteration(PI) and Conservative Policy Iteration(CPI). We then had a very interesting session where we saw model free methods for optimal policy and Temporal Difference Learning. The professor was ideating on one has to think as a ML practitioner in an intuitive way, currently the detailing is something beyond my scope right now. However, sometime in the future I will be able to articulate this aspects better. Later professor presented how RL is actually being applied in Industry by taking a use case from HealthCare domain which was very interesting. We had a session on Attention and Transformers where attention mechanism, evolution from Encoder-Decoder architecture to Transformer architecture and some discussion on state-of-the-art models like GPT(Decoder only) and BERT(Encoder only) models. This was a refresher session for me and helped to understand architectures with much more detailing. Then we had an industry session on how Finance firms like Franklin Templeton are adapting AI for securitized products. This was a very interesting session which helped to dispel the myth that AI will replace everything. This session highlighted on CLO's(Collateral Loan Obligation) and CMBS(Commercial Mortgage-Backend Securities).

Day - 5

We continued on defining Extreme Value Theory (EVT), which focuses on modeling rare but impactful events rather than average behavior. In financial markets, extreme events like sudden crashes, sharp rallies, or volatility spikes often dominate long-term risk, yet they are poorly captured by traditional models that assume normal distributions. I was able to understand EVT has been specifically designed to study the tails of return distributions, where extreme losses and gains reside. Instead of asking, “What happens most of the time?”, EVT asks, “What happens in the worst-case scenarios?” a question that is crucial for risk management in finance.

We explored how EVT helps in:

Estimating the probability of extreme losses
Understanding tail risk beyond variance and standard deviation
Improving risk measures such as Value at Risk (VaR) and CVaR

Unlike standard models that underestimate rare events, EVT provides a framework to model maximum losses over a given time horizon or losses exceeding a high threshold. This makes it especially valuable for stress testing portfolios and evaluating downside risk during market crises.

Another key takeaway was how EVT complements machine learning rather than replacing it. While ML models are often optimized for average performance, EVT-based methods ensure that models remain robust under extreme market conditions. This reinforced the idea that in quantitative finance, risk modeling is as important as return prediction. We had a keynote speaker session by Alpha Grep where we were introduced on how Machine Learning is being used in Quant firms. It was a very interesting session on how there are data requirements for Low, medium and High frequency trading, the data preprocessing challenges and latency issues, how ML still is being used to understand the data. We also touched upon end-to-end data science pipeline and various challenges one has to encompass to get the actual insights from the data. Later we had a session by alpha grep on analyzing HFT data which will involve exploring the data understand each feature come up with alpha vectors and so on. We also had a session on Long and Short LLMs where we walked through different state-of-the-art LLMs and how they can be used in financial firms at operational level to analyze the textual finance data and how there are curated specialized models for the finance domain like FinGPT, Bloomberg GPT and how they are being used by firms o perform sentiment analysis Text classification, NER and Relation extraction.

Day - 6

We had a keynote speaker from Ashoka University where we had a session generating samples from tilted distributions using diffusion models, with a focus on financial applications. This topic connected modern generative AI techniques with classical problems in probabilistic modeling and risk estimation.

I learnt that a tilted distribution arises when we deliberately reweight a base probability distribution to emphasize certain outcomes for example, giving more importance to rare or extreme events. In finance, this is particularly useful because tail events, though infrequent, often have outsized impact on risk and portfolio performance.

Diffusion-based methods provide a powerful framework for sampling from such tilted distributions. Instead of drawing samples directly (can be difficult for complex or high-dimensional distributions) diffusion processes gradually transform noise into structured samples. By incorporating a tilting function into this process, the diffusion can be guided toward regions of interest, such as extreme losses or stress scenarios.

From a financial perspective, this approach has several important applications. It enables more efficient simulation of rare market events, improves estimation of tail-risk measures like Value at Risk and Expected Shortfall, and supports stress testing by focusing computational effort on scenarios that matter most. Then we had a session on Time series modelling. Modelling has a very important role to play in finance domain as the features must be analyzed and transformed to get better insights on data and to design better portfolios as well which we wish to be potentially risk-free. We systematically went through models like AR, MA, ARMA and then ARIMA. Time series data pose challenges like noisy, non-stationary, and often influenced by sudden regime changes. Concepts such as trends, seasonality, volatility clustering, and autocorrelation are essential for deciding which modeling approach is appropriate. Rather than assuming stable patterns, we were encouraged to treat financial markets as evolving systems. A recurring theme was that no single model works universally, model choice depends heavily on the data, time horizon, and objective. Instead of aiming for precise forecasts, time series models are often more useful for risk estimation, scenario analysis, and decision support. This reframed my understanding of modeling success in financial contexts. Another major focus was evaluation and validation. Traditional random train-test splits are inappropriate for time series data, so we learnt the importance of time-aware validation strategies such as rolling windows and walk-forward testing. These techniques better reflect real-world deployment conditions. Then we continued with another session on Risk measures where we studied how VaR and CVaR become important metrics to measure risk and how EVT enables a tradeoff between Bias and Variance in the tail distribution and also introduction to Big theorem that helps estimate risk with less data. Then we had a hands on session on time-series modelling where we practically implemented the AR MA and ARMA models and observed the actual problems which we will have to deal with when it comes to timeseries data.

This marked the end of the Winter School filled with lots of knowledge and experiences.

I sincerely thank all the professors who covered a lot of stuff within 6 days it was a roller coaster and every day I felt excited and was looking forward of what I can learn. This is not the end but just a beginning for me to get into exploring AI and Finance, meaningfully contribute in this field.