qlib/examples/benchmarks/HIST

HIST

• Code: https://github.com/Wentao-Xu/HIST
• Paper: HIST: A Graph-based Framework for Stock Trend Forecasting via Mining Concept-Oriented Shared InformationAdaRNN: Adaptive Learning and Forecasting for Time Series.

Stock-index mapping migration

The bundled qlib_csi300_stock_index.npy object mapping has been replaced by qlib_csi300_stock_index.json, preserving all 735 entries. The example YAML already uses the new file. Update the same setting in your own workflow YAML:

task:
    model:
        kwargs:
            stock_index: "benchmarks/HIST/qlib_csi300_stock_index.json"

This path follows the example's convention of running from examples/; adjust it to your working directory. stock2concept is a separate numeric matrix and still uses a two-dimensional .npy loaded without pickle. Do not rename or replace that matrix with the JSON mapping. It must include the unknown-stock row 733 (at least 734 rows), and all mapped indices must be within its row bounds.

A previously saved HIST model retains its own path. After independently trusting and restoring it, update model.stock_index to the matching JSON file before prediction or further fitting. Changing YAML alone does not update a restored model. Preserve all 735 bundled entries and their original matrix row assignments.

For a known-trusted custom mapping, re-export it from your original trusted metadata or producer into a JSON object:

• Keys must be instrument strings, matching your Qlib dataset.
• Values must be non-negative integers (not strings, floats or booleans) indexing rows in the corresponding stock2concept matrix.
• Preserve each instrument's original row assignment, and verify all indices are within the matrix's row bounds.
• Point task.model.kwargs.stock_index at your new .json file and keep the matching concept matrix.

If you only have an old object-pickled .npy, recover or regenerate the mapping from the trusted source rather than loading an unknown file to convert it. Changing the filename extension alone does not convert the contents. HIST deliberately rejects the old object format because it requires executable pickle deserialization; there is no fallback or recorder trust flag that re-enables it.

See the artifact loading migration guide for the separate policy on model/dataset artifacts. The JSON mapping change does not make pre-existing local model checkpoints or other pickle loaders safe; those inputs still require independent trust.

Full workflow regression

From the repository root, with the test and model dependencies installed:

python -m pytest tests/model/test_hist_workflow.py -m slow -q

This offline CPU regression runs real Alpha360/DatasetH preparation, one HIST training epoch, signal analysis and a six-day TopkDropout backtest in an isolated MLflow store. It checks 48 prediction/label rows, actual optimizer updates, finite reports and trading activity. Saved model/dataset objects are refused by default; explicitly trusted reloads reproduce predictions exactly, including in a fresh Python process. Metadata and concept fixtures are local JSON and numeric NumPy files, with no downloads. This is functional integration coverage, not a paper-scale model-quality benchmark.