Machine Learning Developer in New York, NY, United States

Using various global optimizers, I computed the universe of stock/bond allocations and spending rules to find the plan that would historically have maximized certainty-equivalent spending at various levels of risk aversion. I then created a visualization using React, allowing users to explore optimal rules and the historical spending patterns they would have supported.

Historically, what would have been the optimal allocation and spending plan for retirement?

If you are maximally risk-averse, the Bengen 4% rule holds up: start by spending 4% of your portfolio and increase annually for inflation. If you are less risk-averse, a more flexible rule like a fixed spending amount plus a percentage of assets might make more sense.

Longer writeup: https://www.advisorperspectives.com/articles/2021/03/22/beyond-the-4-rule-flexible-withdrawal-strategies-using-certainty-equivalent-spending

A tutorial on portfolio optimization using mean/variance analysis with cvxpy. For different risk tolerance levels, what is the optimal portfolio allocation between stocks, bonds, cash, and gold? I explored why we do mean-variance analysis, how we can use convex optimization and cvxpy to determine what optimal portfolios would have looked like over various historical periods and whether gold has a place in optimal portfolios.

I built an algorithmic news feed for financial market news. I used the Twitter API, Pagerank, and NLP to rank the most influential accounts focused on financial markets and aggregate the news stories they shared. The public version is offline, but a demo is available on request.

An overview of machine learning classification methods and application to a factor investing problem by attempting to forecast future performance quantiles. While seemingly quite successful, the result seems to demonstrate overfitting, and regression and learn-to-rank outperform vanilla multinomial classification with softmax loss.

I wrote a blog post exploring analogies between investing and playing the game of poker. Both are games of quantitative thinking about imperfect information and psychology. But real-world markets don't reward bluffers.

This is a thought experiment on what would happen if more and more people invest passively. There's no such thing as a 100% purely passive investor: we are free to pick a strategy but not avoiding the choice. As more people invest more passively, they save on expenses, but market efficiency decreases. The ideal "Warren Buffett" active investor can always exploit any "suckers." At some point, though, the less active investing there is, the more excellent opportunities there are for remaining active investors to exploit.

An overview of deep reinforcement learning, code examples with OpenAI gym Cartpole and LunarLander, and code to implement policy gradient on market data, with positive results on simulated data (Ornstein–Uhlenbeck mean reversion process and noise).

Here on a blog post, I had fun with topic modeling, clustering, Word2vec, and text generation with fund names. A recurrent neural network trained on a corpus of fund names generates new fund names based on any starting letters.

Used Mathematica to explore optimal bluffing frequencies and call frequencies in poker based on simplified assumptions. As long as you have a positive EV to bluffing, you should increase your bluff frequency. There is a Nash equilibrium where the EV to bluffing more frequently and calling more frequently are 0.

A platform that allows users to search on Google Maps, Yelp, and Foursquare for a keyword like 'Pizza' displays a combined Bayesian ranking and results on a map.

Back-end stack: Python, Flask, deployed as a Docker microservice in Amazon ECS.

Front-end stack: Bootstrap, jQuery, Leaflet.js.