Projects

Repositories are on my Github

• I modeled crossword navigation as a stochastic shortest path problem and used dynamic programming to study how clue-ordering strategies affect solve time. Using data from my own solving behavior, I showed that even simple heuristic-based strategies reduce expected solve time by nearly 20% compared to baseline approaches. Most of the benefit comes from shallow lookahead that accounts for how information propagates through the grid.

• I wrote a paper exploring the theory of hidden Markov models and their applications in quantitative finance. I wrote Python code to implement the Baum-Welch algorithm for 2-state Bernoulli hidden Markov models, and used a bootstrapping technique to estimate the expected Fisher information matrix for latent processes with various parameters. I fit a 2-state Gaussian HMM to real-world S & P 500 data and analyzed its potential in predicting future market volatility.

• I conducted a study of teaser bets using 25 years of NFL data. By applying logistic regression, I identified conditions under which teaser legs could yield positive expected returns.

• I wrote a follow-up article that proposes an original, market-neutral strategy for finding positive expected value teaser bets using arbitrage.

• I built a real-time win probability model trained on play-by-play data using XGBoost and neural nets and analyzed its performance. Use my model with the app below.