CASE 05 · FINANCIAL MACHINE LEARNING

QRT Allocation Performance Prediction

Panel data · Chronological validation · Explainable weak-signal modeling

TRAINING PANEL527,073

Allocation-time observations.

CROSS SECTION278

Unique financial allocations.

TIME AXIS2,522

Anonymized timestamps.

ENGINEERED INPUTS73

Raw and derived features before clustering.

RESEARCH DESIGN

Weak signals require strict validation.

The objective was to predict the sign of a future allocation return. Every preprocessing step was fitted on the past before evaluation on a later validation block.

Audit

Measured balance, missingness, allocation coverage and time dependence.

Engineer

Built 27 momentum, volatility, shape, trend and liquidity features.

Split

Trained on earlier timestamps and validated only on later observations.

Compare

Benchmarked logistic regression, random forest, LightGBM and dense MLP models.

Explain

Audited feature importance and SHAP effects before final submission.

EMPIRICAL EVIDENCE

A small edge, reported without inflation.

Random Forest produced the highest raw validation accuracy, while LightGBM was retained for the final submission because it preserved nearly the same score with a materially smaller train-validation gap.

Validation accuracy comparison across four financial allocation models — FIGURE 01 · MODEL SELECTION51.90% LR · 52.46% RF · 52.39% LightGBM · 52.10% MLP
All scores use the same chronological validation split; the dotted reference shows the majority-class baseline.

Random Forest and LightGBM financial feature importance — FIGURE 02 · IMPORTANCEShort-horizon returns and turnover dominate
RET_1, RET_2, RET_8 and median daily turnover repeatedly rank among the strongest inputs.

SHAP summary for the LightGBM allocation model — FIGURE 03 · ATTRIBUTIONSHAP direction and magnitude
The explanation layer makes nonlinear effects visible instead of treating a small accuracy gain as a black box.

MODEL CHOICE

Stability over leaderboard chasing.

LightGBM reached 52.39% accuracy and 0.5313 AUC with a +0.0431 train-validation gap, versus +0.0837 for the slightly higher-scoring Random Forest.

LIMIT

Predictive, not yet tradable.

The challenge metric is classification accuracy. Turnover, execution costs, portfolio construction and economic capacity remain outside this evaluation.