From 4bbd990b01700f4c7e8c0432cdec71a7b0a5dede Mon Sep 17 00:00:00 2001 From: =?utf8?q?J=C3=A9r=C3=B4me=20Benoit?= Date: Thu, 17 Sep 2026 23:00:42 +0200 Subject: [PATCH] docs(quickadapter): remove operational constraints section --- README.md | 49 ------------------------------------------------- 1 file changed, 49 deletions(-) diff --git a/README.md b/README.md index 8866b4d..d3b52cd 100644 --- a/README.md +++ b/README.md @@ -169,55 +169,6 @@ ignored with a warning. Matching column patterns are applied from least to most specific; equally specific patterns follow declaration order, so the later one wins. -Operational constraints: - -- Entries require every requested prior confirmation candle. Exits may use the - available history, but still reject a failed measurable comparison. Partial - take-profit exits check both the outgoing quantity and the remainder against - exchange minimums after contract precision rounding, including leverage. - An infeasible partial exit requests a full exit; this cannot guarantee that an - exchange accepts an already-dust position. -- `freqai.fit_live_predictions_candles` defaults to 100. Missing or invalid - values normalize to 100 in the shared strategy/FreqAI configuration; valid - positive integers are preserved. -- Label HPO scheduling permits pairs to share slots when there are fewer slots - than pairs. Scheduling is inactive when HPO is disabled. -- Causal training removes unavailable terminal labels even with `test_size=0`. - Savitzky–Golay `interp` edge availability includes the whole edge-fit window - for labels and weights. A window longer than the series leaves both unchanged. -- Cluster ranking and member selection honor supported objective weights. KNN - uses weighted SciPy distances, excludes only each row's own index, and retains - distinct candidates at zero distance. Its distance matrix requires quadratic - memory in the number of Pareto candidates. -- Standardized Euclidean variance and Mahalanobis covariance are estimated once - from the full normalized Pareto front, after constant objectives are removed. - Singular covariance eigenvalues are floored at the largest eigenvalue times - the square root of machine epsilon. These two metrics do not support custom - objective weights; configured weights warn and fall back to uniform weights. -- Objective weights are validated once and normalized by maximum then sum; - scale-invariant selection survives any positive rescaling of the same - weights. An all-zero configuration falls back to uniform weights. -- Combined-metric softmax aggregation is numerically stable for any finite - positive temperature and coefficients; log coefficients are applied after - temperature division. Boolean metric coefficients are ignored. -- Invalid label metric values of any JSON type follow the documented - warning/fallback policy instead of raising. -- Reversed label HPO bounds publish their validated fallback to the shared - configuration, so the label objective receives consistent ranges. -- With `space_reduction=true`, `space_fraction=0` fixes numerical search ranges - at valid previous best values; it does not freeze every categorical choice. - Saved HPO suggestions reconstruct derived estimator parameters for final fits: - XGBoost `lossguide` uses unlimited depth and histogram gradient boosting - restores the selected zero-regularization branch. -- NGBoost defaults to `model_training_parameters.dist="normal"`, which supports - signed labels. Explicit `"lognormal"` requires finite, strictly positive - transformed training and validation labels; incompatible HPO trials are pruned, - while direct fits raise an error. -- Changing objective-weight normalization, softmax stabilization, metric-type - validation, or bound publication resets label HPO selection state - (selection schema 3). The persisted best-params JSON layout remains - independently versioned. - ### Backtest evaluation protocol Use this protocol before adopting a change to a QuickAdapter default. It is an -- 2.53.0