Logistic Signal Calibration

Logistic Signal Calibration implements logistic signal calibration as a transparent on-chart model: momentum, trend and volume-flow features are standardized to z-scores, weighted by coefficients set directly in the inputs, summed with an intercept into log-odds, and mapped through the sigmoid into a 0-to-1 score. The pane plots the calibrated probability against decision thresholds, and a reliability dashboard grades every score once its horizon resolves.

How to Trade the Logistic Signal Calibration?

• Score above the Bullish threshold: the long filter passes - marginal readings near 0.5 are skipped by construction.
• Score below the Bearish threshold: the short mirror; between the two is the no-edge, stand-aside band.
• Overconfidence highlights: bins where claimed probability exceeds observed frequency beyond one standard error turn orange.
• Brier score above 0.25: worse than a constant always-0.5 forecast; a calibration-decay alert fires on the cross.

Per-feature contributions stream to the data window, keeping every reading decomposable.

Logistic Signal Calibration Settings

• Source (default close).
• Momentum (default 14), Trend (default 50), Volume (default 20): the feature lookbacks, each paired with its logistic coefficient (default 1; 0 removes a feature, negative flips it).
• Intercept (Bias) (default 0): baseline log-odds - 0 rests the score at 0.5.
• Normalization Window (default 100): trailing z-score window for the features.
• Bullish (default 0.7) / Bearish (default 0.3): decision thresholds.
• Prediction Horizon (default 10): bars until each score resolves.
• Evaluation Window (default 200): resolved scores in the rolling calibration sample.
• Show Calibration Dashboard (default on); style toggles cover gradient fill, cross markers and line width.

Frequently Asked Questions

Are the scores real win probabilities?

Treat them as ranked confidence: calibration is measured on past data and erodes as conditions drift. The hit-rate-versus-confidence row and the Brier baseline continuously test whether the numbers still mean what they claim.

How does this differ from a Bayesian classifier?

Bayesian Classifiers multiply per-feature likelihoods and grow overconfident when features overlap; the logistic form combines features additively in log-odds and stays auditable input by input.

Why are the weights manual rather than fitted?

Transparency: fitted coefficients come from regression on labeled history and change with every refit. Setting them directly keeps cause and effect visible - and the dashboard still reports honestly how the chosen weights perform.