Temporal Routing Adaptor for Learning Multiple Stock Trading Patterns

# Learning Multiple Stock Trading Patterns with Temporal Routing Adaptor and Optimal Transport


# Learning Multiple Stock Trading Patterns with Temporal Routing Adaptor and Optimal Transport

Temporal Routing Adaptor (TRA) is designed to capture multiple trading patterns in the stock market data. Please refer to [our paper](http://arxiv.org/abs/2106.12950) for more details.

If you find our work useful in your research, please cite:
```
@inproceedings{HengxuKDD2021,
 author = {Hengxu Lin and Dong Zhou and Weiqing Liu and Jiang Bian},
 title = {Learning Multiple Stock Trading Patterns with Temporal Routing Adaptor and Optimal Transport},
 booktitle = {Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery \& Data Mining},
 series = {KDD '21},
 year = {2021},
 publisher = {ACM},
}

@article{yang2020qlib,
  title={Qlib: An AI-oriented Quantitative Investment Platform},
  author={Yang, Xiao and Liu, Weiqing and Zhou, Dong and Bian, Jiang and Liu, Tie-Yan},
  journal={arXiv preprint arXiv:2009.11189},
  year={2020}
}
```

## Usage (Recommended)

When upgrading, consult the [configuration migration guide](../../../docs/start/config_migration.rst).
It covers the unreleased source changes, including explicit `MODEL_TYPES`
registration for custom built-in TRA backbones. Register extensions before model
construction in each worker; source/main/tagged installations may differ.

**Update**: `TRA` has been moved to `qlib.contrib.model.pytorch_tra` to support other `Qlib` components like  `qlib.workflow` and `Alpha158/Alpha360` dataset.

Please follow the official [doc](https://qlib.readthedocs.io/en/latest/component/workflow.html) to use `TRA` with `workflow`. Here we also provide several example config files:

- `workflow_config_tra_Alpha360.yaml`: running `TRA` with `Alpha360` dataset
- `workflow_config_tra_Alpha158.yaml`: running `TRA` with `Alpha158` dataset (with feature subsampling)
- `workflow_config_tra_Alpha158_full.yaml`: running `TRA` with `Alpha158` dataset (without feature subsampling)

The performances of `TRA` are reported in [Benchmarks](https://github.com/microsoft/qlib/tree/main/examples/benchmarks).

## Usage (Not Maintained)

This section is used to reproduce the results in the paper.

### Running

We attach our running scripts for the paper in `run.sh`.

Run the commands below from `examples/benchmarks/TRA`. The legacy configurations
load `src/model.py` and `src/dataset.py`, so **both** `task.model` and
`task.dataset` declare top-level `trusted: true`, alongside `class`,
`module_path`, and `kwargs`. Relative file paths use the current working
directory, not the YAML file. Only use these settings with code and
configurations you trust; consent is not inherited from another component.

If you copied an older configuration, add these per-component booleans and
remove any `qlib_init.trusted_module_roots` setting from earlier PR revisions.
If you copied `example.py`, retain full initialization forwarding with
`qlib.init(**config["qlib_init"])`; it forwards settings, not file-import
permission. See the [migration guide](../../../docs/start/config_migration.rst)
for strict boolean semantics, workers, and independent artifact permissions.
The legacy `model_type: LSTM` belongs to `src/model.py`; do not rename it just
because the separate built-in TRA implementation supports `RNN` and `Transformer`.

And here are two ways to run the model:

* Running from scripts with default parameters

  You can directly run from Qlib command `qrun`:
  ```
  qrun configs/config_alstm.yaml
  ```

* Running from code with self-defined parameters

  Setting different parameters is also allowed. See codes in `example.py`:
  ```
  python example.py --config_file configs/config_alstm.yaml
  ```

Here we trained TRA on a pretrained backbone model. Therefore we run `*_init.yaml` before TRA's scripts.

### Results

After running the scripts, you can find result files in path `./output`:

* `info.json` - config settings and result metrics.
* `log.csv` - running logs.
* `model.bin` - the model parameter dictionary.
* `pred.pkl` - the prediction scores and output for inference.

Evaluation metrics reported in the paper:
This result is generated by qlib==0.7.1.

| Methods | MSE| MAE| IC | ICIR | AR | AV | SR | MDD |
|-------|-------|------|-----|-----|-----|-----|-----|-----|
|Linear|0.163|0.327|0.020|0.132|-3.2%|16.8%|-0.191|32.1%|
|LightGBM|0.160(0.000)|0.323(0.000)|0.041|0.292|7.8%|15.5%|0.503|25.7%|
|MLP|0.160(0.002)|0.323(0.003)|0.037|0.273|3.7%|15.3%|0.264|26.2%|
|SFM|0.159(0.001)	|0.321(0.001)	|0.047	|0.381	|7.1%	|14.3%	|0.497	|22.9%|
|ALSTM|0.158(0.001)	|0.320(0.001)	|0.053	|0.419	|12.3%	|13.7%	|0.897	|20.2%|
|Trans.|0.158(0.001)	|0.322(0.001)	|0.051	|0.400	|14.5%	|14.2%	|1.028	|22.5%|
|ALSTM+TS|0.160(0.002)	|0.321(0.002)	|0.039	|0.291	|6.7%	|14.6%	|0.480|22.3%|
|Trans.+TS|0.160(0.004)	|0.324(0.005)	|0.037	|0.278	|10.4%	|14.7%	|0.722	|23.7%|
|ALSTM+TRA(Ours)|0.157(0.000)	|0.318(0.000)	|0.059	|0.460	|12.4%	|14.0%	|0.885	|20.4%|
|Trans.+TRA(Ours)|0.157(0.000)	|0.320(0.000)	|0.056	|0.442	|16.1%	|14.2%	|1.133	|23.1%|

A more detailed demo for our experiment results in the paper can be found in `Report.ipynb`.

## Common Issues

For help or issues using TRA, please submit a GitHub issue.

Sometimes we might encounter situation where the loss is `NaN`, please check the `epsilon` parameter in the sinkhorn algorithm, adjusting the `epsilon` according to input's scale is important.