AlphaSeeker — TradeMaster 2026

Financial time-series forecasting artifact from the HKUST(GZ) TradeMaster Cup 2026, where our team placed 5th. This repository preserves the modeling work behind the competition submission and makes the evaluation assumptions explicit enough to distinguish an offline forecasting artifact from a deployable trading system.

Scope

The task uses minute-indexed tabular features and three forecasting targets with different horizons. The repository currently contains three modeling paths:

Path	File	Role
XGBoost	models/submissionpipelinexgb.py	tree baseline with daily normalization, interaction features, chronological train/validation split, and weighted MAE evaluation
GRU	models/final_gru.py	30-step sequence model with contextual daily statistics and weighted multi-horizon loss
Blend	models/blend.py	fixed 50/50 submission-level combination of model outputs

The code is retained close to the competition implementation rather than rewritten to imply a cleaner experiment than was actually run. Comments and output names have been normalized where the original competition-time wording overstated leakage protection or carried unrelated temporary labels.

Evaluation contract

The competition objective combines three target horizons using a weighted MAE:

[ L = 0.1,\mathrm{MAE}{short} + 0.3,\mathrm{MAE}{medium} + 0.6,\mathrm{MAE}_{long}. ]

Both main pipelines split training and validation data by date_id, so the validation period is temporally later than the training period. This is preferable to a random row split for non-stationary financial data.

That chronological split does not by itself make every feature causal. The information-set boundary is discussed below.

Information-set and leakage audit

The current competition code performs full-day normalization within each date_id.

• submissionpipelinexgb.py computes each feature's daily mean and standard deviation before constructing daily z-scores.
• final_gru.py concatenates train and test frames for feature construction and also computes full-day means/standard deviations within each day.

For an offline competition where the complete daily feature matrix is supplied, this can be a valid feature transformation under the competition's data-access contract. For a live minute-by-minute forecasting interpretation, the same transformation would use observations from later minutes of the day and therefore violate the causal information set.

A research or deployment version should replace full-day statistics with one of the following, depending on the intended decision time:

1. lagged statistics estimated from previous days only;
2. expanding intraday statistics using observations available up to time t;
3. rolling statistics with an explicitly lagged window;
4. cross-sectional normalization at a timestamp when the full cross-section is genuinely observable.

Any future result reported from this repository should state which information set is used. Competition scores and causal/out-of-time research results should not be mixed.

Reproducibility boundary

The original competition dataset is not redistributed here. The scripts expect:

dataset/
├── train_v2.csv
└── test_v2.csv

with at least:

• id
• date_id
• minute_id
• feature_1 ... feature_30
• target_short, target_medium, target_long in the training set

Install the preserved Python dependencies, place the competition data under dataset/, then run the paths from the repository root:

python models/submissionpipelinexgb.py   # -> submission_xgb.csv
python models/final_gru.py               # -> submission_gru.csv
python models/blend.py                   # -> submission_blend.csv

These are research snapshots rather than a packaged library. Model training, validation, inference, and submission generation remain in the original standalone pipelines so the competition path stays auditable. CI therefore performs syntax/research-contract checks without pretending to reproduce scores in the absence of the private competition dataset.

What this repository demonstrates

The useful part of this project is the modeling and evaluation process under noisy financial time series:

• chronological validation instead of random row splitting;
• comparison of tree-based and recurrent models;
• multi-horizon loss weighting;
• feature normalization and interaction design;
• explicit recognition that offline leaderboard features and deployable causal features have different information constraints.

The next research-quality revision should add a leakage-safe feature pipeline and walk-forward evaluation before making any claim about out-of-sample trading usefulness.

Status

Archived competition artifact / research baseline. The repository is being cleaned up for reproducibility and methodological clarity. It should not be interpreted as investment advice or as evidence of a production trading strategy.