Comparing Online Logistic Regression Update Rules on Candlestick Data

Summary

This research strategy compares four logistic regression approaches on the same candlestick stream: a fixed model, per-bar online updates, periodic retraining on a recent rolling window, and updates triggered by worsening prediction loss. It uses price and volume-derived features, scales them, predicts the next bar’s direction, and updates models only after evaluating the preceding prediction. The strategy keeps four shadow portfolios and reports predictive and trading diagnostics, including log loss, accuracy, virtual return, drawdown, exposure, turnover, direction changes, and parameter movement.

The description emphasizes controlled comparison: all models share data, features, initial training, signal thresholds, and cost assumptions. It avoids treating an unfinished bar as complete and evaluates virtual returns from the following bar’s open to close. These safeguards improve consistency, but the excerpt supplies no comparative results or evidence that any update rule performs better. The virtual accounts do not model real order execution, and the code is presented as a research experiment rather than a live trading system.

Key ideas

• The experiment compares fixed, continuously updated, periodically retrained, and loss-gated logistic regression models.
• Each model uses common features, initialization, thresholds, and assumed trading costs to make the comparison more controlled.
• The system evaluates prior predictions before training on newly observed labels.
• A rolling window supports periodic retraining and monitors recent loss for gated updates.
• Shadow-account returns are simulated and do not represent actual exchange execution.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.