Comparing Online Model Update Speeds for Trading Signals
Summary
This controlled study asks whether frequent model updates adapt to changing markets or absorb short-term noise. It compares four logistic regression variants in one Rust process: a fixed model, per-sample online updates, periodic rolling-window retraining, and updates gated by deteriorating recent Log Loss. They share initialization, features, standardization, trading thresholds, virtual execution rules, and cost assumptions. The setup predicts the next candle’s open-to-close direction from information available at the prior close, scores the prediction before training on its label, and tracks predictive and trading measures including Log Loss, turnover, drawdown, and parameter movement.
In the reported default backtest, all four models lose heavily after costs; more frequent updating does not improve out-of-sample Log Loss. Gated updating reduces turnover and direction flips but does not produce a trading advantage. The study cautions that one asset, interval, and backtest cannot establish general results. Virtual fills, fixed costs, fixed feature scaling, and heuristic gating limit interpretation, and the near-zero net values make some return comparisons uninformative. It recommends cost scenarios, adjacent parameter checks, and comparisons across markets and regimes.
Key ideas
- The experiment isolates update speed by holding model structure, features, initialization, and trading rules constant.
- The four update schemes are fixed parameters, per-sample online learning, periodic rolling retraining, and loss-gated updates.
- Predictions are evaluated before the corresponding labels are used for training, reducing leakage from the current sample.
- Frequent updates did not improve out-of-sample Log Loss in the reported run, while gating reduced turnover without creating a trading edge.
- Virtual fills, simplified costs, fixed scaling, and a single test setup limit how broadly the results can be applied.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.