Model Selection Guide

Language: English · 简体中文

Status: active
Owner: quant-market-research
Last verified: 2026-07-16
Source of truth: yes
Superseded by: n/a

Use this guide to choose a starting point among the trainable models. The registered model set and extension criteria are documented in Model Landscape. Parameter and command details are in orchestration configuration and CLI helpers.

The registry also includes fixed_score_artifact. It replays frozen external scores without training a prediction model, so it belongs to the artifact-replay workflow.

Quick choices

Research goal	Starting model
Establish a nonlinear baseline	xgb_regressor
Test whether a same-day ranking objective adds value	xgb_ranker
Check a linear signal with mixed-scale features	ridge_scaled
Use a compatibility baseline for already aligned features	ridge
Test sparse linear constraints	elasticnet
Replay frozen external scores	fixed_score_artifact

Model roles

Model	Training approach	Strength	Limitation
xgb_regressor	Fits a continuous target	Models nonlinear relationships and feature interactions	Requires stable time splits and checks for degenerate predictions
xgb_ranker	Learns rankings grouped by trade_date	Its objective matches cross-sectional ranking	Requires valid within-date groups and costs more to train and tune
ridge	Linear regression with L2 regularization	Fast, interpretable coefficients, and relatively stable under collinearity	Cannot represent complex nonlinear relationships
ridge_scaled	Fits scaling within each training fold, then applies Ridge	Handles financial features with mixed units without letting scale dominate regularization	Still linear; scaling must never be fit on test folds or future dates
random_forest_regressor	Random-forest regression	Provides a nonlinear tree-averaging baseline	Still needs time splits and mature labels; out-of-bag scores do not replace out-of-time validation
elasticnet	Linear regression with L1 and L2 regularization	Shrinks coefficients and can remove features	Requires choosing both alpha and l1_ratio; all-zero predictions are more likely

The A-share preset currently uses xgb_regressor. For financial research, start with ridge_scaled as the linear comparator and use ridge as a compatibility check. Neither replaces strict time-based validation.

Minimal configurations

Ridge

model:
  type: ridge
  params:
    alpha: 1.0
  sample_weight_mode: date_equal

Use ridge to check whether features have a stable linear relationship with a continuous target. If the linear result is weak, first review the data, target, and feature evidence before adding model complexity.

Scaled Ridge

model:
  type: ridge_scaled
  params:
    alpha: 1.0

ridge_scaled fits StandardScaler on each training fold before fitting Ridge. This is useful when financial features mix percentages, amounts, and ratios. Never fit the scaler on a test fold or future dates.

XGBoost ranker

model:
  type: xgb_ranker
  params:
    objective: rank:pairwise

xgb_ranker groups training rows by trade_date. After establishing a reproducible regression baseline, use it to test whether directly optimizing ranking improves cross-sectional results.

ElasticNet

model:
  type: elasticnet
  params:
    alpha: 1.0
    l1_ratio: 0.5
  sample_weight_mode: date_equal

Use elasticnet as a focused sparse-linear comparison. After a run, inspect flag_constant_prediction and flag_zero_feature_importance in summary.json.

Frozen-score artifact

model:
  type: fixed_score_artifact
  params:
    score_col: pred

The input must contain the column named by score_col. This path replays frozen scores as supplied and can validate an external strategy or a cross-repository signal handoff.

Search discipline

The current public strategy-pipeline CLI does not expose alpha model-tuning or linear-sweep commands. The historical strategy alpha tune, strategy alpha sweep-linear, and strategy summarize examples are not supported by the current CLI. Confirm available commands with strategy-pipeline --help before relying on older notes.

When conducting a model search through a supported research runner, keep these boundaries:

1. Fix the data, target, features, and time split first.
2. Search model.params, sample_weight, and train_window during training.
3. After selecting a signal, use strategy backtest grid to test top_k, costs, buffers, and weights.
4. For small samples, use min_cv_ic_valid_folds to exclude trials with too few valid folds.
5. Keep the full trial ledger; do not record only the best configuration.

Checks after a run

Inspect the run summary and evaluation artifacts for:

1. eval.constant_prediction for constant predictions.
2. eval.zero_feature_importance for zero feature importance.
3. train_ic and test_ic for the training/test gap.
4. backtest_sharpe, turnover, and costs for execution relevance.
5. Walk-forward and final out-of-sample evidence for stability over time.

For a fair model comparison, hold data assets, targets, time splits, costs, and portfolio construction constant. See Overfitting Controls for research governance.