Blog

Flash Attention makes a 23,400-step trading context computationally free — tiling, online softmax, and an IO bound of N^2 d^2 / M. Whether that longer context makes the model better is a separate question, and it has to be measured, not asserted.

Three things about trading ensembles this blog has not covered: why the covariance term — not model count — dominates ensemble error, how stacking works as a signal-combination layer with out-of-fold meta-features, and whether turnover netting across many signals actually reduces execution cost.

Every model on this blog so far answers 'what predicts what?'. Double ML answers 'what causes what?' — with a standard error you can defend. The partially linear model, Neyman orthogonality, purged cross-fitting on order book data, and an honest account of why a valid DML confidence interval survives exactly one pre-specified question.

Every backtest on this blog estimates a conditional mean. Causal forests estimate a conditional treatment effect instead — tau(x) rather than mu(x) — with honest splitting, an adaptive-kernel weight representation, and a calibration test that tells you whether the heterogeneity you found is real.

Every sizing rule on this blog treats uncertainty as one number. MC Dropout and deep ensembles split it into model ignorance and market noise — and those two deserve different position sizes.

Automated feature generation (tsfresh, Featuretools), budget-aware model search (FLAML's cost-frugal optimizer), and the WorldQuant formulaic alpha factory — the parts of the research pipeline this blog's search-and-overfit arc never covered, and what the arc's own results say about them.

Every leveraged position advertises the price at which it must be sold. How to build the on-chain liquidation depth chart and the CEX liquidation heatmap, model cascade dynamics as a reproduction number, and trade forced flow as a signal instead of only fearing it as a risk.