Essays on rigorous backtesting, statistical methodology, and the realities of retail algorithmic trading. No "guaranteed signals", no hype — just the maths.
A backtest can be wrong in three distinct ways. Without proper statistics you'll find a strategy that “works” on every dataset. This post walks through multiple-comparison bias, point-estimate hubris, and in-sample fitting — and shows how each one quietly inflates the strategies you take to live.
Read the post →A single Sharpe number with no confidence interval is a coin flip in disguise. We walk through why, with a concrete example where the 95% CI on the Sharpe runs from -0.18 to +1.42 — making the headline "0.74" almost meaningless. Plus what to report instead.
Read the post →k-fold cross-validation is the default in ML, but it leaks information in time-series in a quiet, technical way that inflates apparent edge by anywhere from 20% to 200%. Why walk-forward is the right primitive for financial backtests — and what it explicitly does and doesn't prove.
Read the post →