Applications
Small, self-contained demos of analytical methods applied to financial and statistical problems. All use public data or synthetic DGPs; code is reproducible; there is no Amazon content here.
These are deliberately compact, each one targets a single technique, shows a single representative figure, and documents the method at the level of a short applied-stats primer. Several originate from long-running interests predating my time at Amazon: my Citi tenure in MBS analysis, my Insight / Georgetown work on financial time-series forecasting, my Quantopian-era statistical-arbitrage research.
The applications
1. Stock-price forecasting with an LSTM
Sequence-to-sequence forecasting of equity prices. The honest story: LSTMs on price series do better than random-walk only under narrow conditions, and their performance on log-returns (the right target) is rarely distinguishable from simpler baselines. Includes a synthetic-but-realistic return series demonstrating the forecasting pipeline and its residual diagnostics.
2. Pairs trading and cointegration
Two assets that share a stochastic trend but diverge around it can be traded when their spread deviates far from its mean. The statistical foundation is cointegration (Engle-Granger 1987; Johansen 1988). Demo shows the two nonstationary series, the stationary spread, and rule-based entry/exit signals.
3. Portfolio optimization and the efficient frontier
Markowitz (1952) gave the first analytical framework for choosing portfolio weights given asset return means and covariances. The demo constructs the efficient frontier for a 5-asset universe, identifies the max-Sharpe portfolio, and discusses the practical failures of the approach (parameter estimation error dominates in practice).
4. Mortgage prepayment modeling
The central modeling object in mortgage-backed-securities pricing is the conditional prepayment rate (CPR) as a function of borrower incentive. The demo walks through the S-curve parameterization, the “burnout” effect (cohorts that didn’t refinance when it made sense become less responsive to future opportunities), and the implications for MBS cash-flow pricing.
What these applications are NOT
- Not production trading systems. They are pedagogical demos.
- Not backed by proprietary data. All figures use synthetic or public datasets.
- Not related to my Amazon work. The applications live here because they reflect long-term quantitative interests that predate and are independent of my current role.
Reproducibility
Figure generation scripts for all four applications are available in the site repository. Each script runs in a few seconds with standard scientific Python (numpy, matplotlib, scikit-learn).