A political-economy PhD with Python and public stock-price data tries to reproduce a small piece of quantitative trading by building three strategies: a one-year momentum model, pairs trading that bets on correlated stocks’ prices converging, and a Ridge-regression model that ranks roughly 45 stocks using momentum, trend, and volatility signals. In the backtests, momentum returned about 10% annually, pairs trading made mild losses, and the machine-learning strategy rose more than sevenfold from 2017, with a reported 25.7% annual return and a Sharpe ratio of 0.96. That performance came with the largest drawdowns. The comparison is not proof of genuine skill: the strategies have different market exposures, and returns attributed to “alpha” may instead reflect familiar factors such as momentum or large-company exposure.
The experiment illustrates both how accessible basic quant methods have become and why professional results are difficult to reproduce. Its assumptions - clean, survivorship-free data, tiny trading costs, perfect fills, unlimited capacity, stable patterns, and no market shocks - are far more favorable than real trading conditions, while a flexible model can overfit historical returns. Professional firms have advantages the laptop test lacks: detailed order-book and alternative data, many specialized models combined across time horizons, and sophisticated execution and risk systems that control market impact and exposure. The central lesson is that coding a plausible strategy is relatively easy; establishing that it has durable, risk-adjusted edge after realistic costs and changing market conditions is the hard part.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.