Unit 13 of 14
Statistical quant research
- 8 lessons
- Videos 97 to 104
Quantitative research asks whether a signal predicts returns after costs. This unit sets out the information, validation and testing rules that keep such a study honest, and the limits that costs set.
Point-in-time data records what was known at each date, and survivor selection or corporate actions can change measured returns. A prediction target and its features are defined relative to an information set and a horizon. Regularization, calibration and distribution shift are separate questions. Validation follows time, and overlapping labels are purged. Searching many strategies calls for Bonferroni or false discovery rate control. Cointegration describes a stationary spread, Sharpe ratios need care with serial dependence, and costs limit capacity. No check guarantees future performance.
Every lesson will be one video. The videos are coming soon, and each lesson page already has the definitions and an example for every key term.
Data, prediction and learning
Point in time financial data and sampling
How timestamps and an information set fix what a past decision could know, how point-in-time data handle data revisions, and how survivorship bias and corporate action adjustment change measured returns.
Video 97
Financial prediction and feature construction
How a prediction target, its label horizon and admissible features with known availability are defined relative to an information set, and why the conditional mean predictor is optimal for squared-error prediction.
Video 98
Machine learning losses regularization and calibration
How empirical risk and the ridge estimator define a regularized fit, why the Brier loss is a proper loss, what probability calibration means, and when covariate shift lets reweighting recover risk under distribution shift.
Video 99
Validation, multiple testing and cointegration
Time series validation and overlapping labels
How chronological splits and walk-forward validation follow the forecast timeline, when overlapping labels share increments, how purging and an embargo remove them, and how nested validation keeps tuning apart from evaluation.
Video 100
Multiple testing and backtest overfitting
How searching many strategies changes false-rejection probabilities, how Bonferroni controls the family-wise error rate, how Benjamini-Hochberg controls the false discovery rate, and what backtest overfitting means.
Video 101
Cointegration and statistical arbitrage
What integration order and cointegration mean, how a common trend cancels in a stationary spread, how an autoregressive spread gives an error-correction relation, and why statistical arbitrage is not arbitrage.
Video 102
Signals, costs and capacity
Signal evaluation and performance attribution
How the information coefficient, a strategy return and the Sharpe ratio differ, how serial dependence enters Sharpe ratio aggregation over several periods, and how factor attribution splits a mean return.
Video 103
Turnover costs and strategy capacity
How turnover and an impact curve set the trading cost and net return of a strategy, and why its break-even capacity differs from its profit-maximizing capital under a stated quadratic cost law.
Video 104