This page lists a batch of methods designed for alpha seeking. Each method tries to give scores/predictions for all stocks each day(e.g. forecasting the future excess return of stocks). The scores/predictions of the models will be used as the mined alpha. Investing in stocks with higher scores is expected to yield more profit.

The alpha is evaluated in two ways.

1. The correlation between the alpha and future return.
2. Constructing portfolio based on the alpha and evaluating the final total return.
   • The explanation of metrics can be found here

Here are the results of each benchmark model running on Qlib's Alpha360 and Alpha158 dataset with China's A shared-stock & CSI300 data respectively. The values of each metric are the mean and std calculated based on 20 runs with different random seeds.

The numbers shown below demonstrate the performance of the entire workflow of each model. We will update the workflow as well as models in the near future for better results.

NOTE: The backtest start from 0.8.0 is quite different from previous version. Please check out the changelog for the difference.

NOTE: We have very limited resources to implement and finetune the models. We tried our best effort to fairly compare these models. But some models may have greater potential than what it looks like in the table below. Your contribution is highly welcomed to explore their potential.

• The selected 20 features are based on the feature importance of a lightgbm-based model.
• The base model of DoubleEnsemble is LGBM.
• The base model of TCTS is GRU.
• About the datasets
  • Alpha158 is a tabular dataset. There are less spatial relationships between different features. Each feature are carefully designed by human (a.k.a feature engineering)
  • Alpha360 contains raw price and volue data without much feature engineering. There are strong strong spatial relationships between the features in the time dimension.
• The metrics can be categorized into two
  • Signal-based evaluation: IC, ICIR, Rank IC, Rank ICIR
    • 
    • 
    • 
    • 
    • 
  • Portfolio-based metrics: Annualized Return, Information Ratio, Max Drawdown

The results on CSI500 is not complete. PR's for models on csi500 are welcome!

Transfer previous models in CSI300 to CSI500 is quite easy. You can try models with just a few commands below.

cd examples/benchmarks/LightGBM
pip install -r requirements.txt

# create new config and set the benchmark to csi500
cp workflow_config_lightgbm_Alpha158.yaml workflow_config_lightgbm_Alpha158_csi500.yaml
sed -i "s/csi300/csi500/g"  workflow_config_lightgbm_Alpha158_csi500.yaml
sed -i "s/SH000300/SH000905/g"  workflow_config_lightgbm_Alpha158_csi500.yaml

# you can either run the model once
qrun workflow_config_lightgbm_Alpha158_csi500.yaml

# or run it for multiple times automatically and get the summarized results.
cd  ../../
python run_all_model.py run 3 lightgbm Alpha158 csi500  # for models with randomness.  please run it for 20 times.

Your contributions to new models are highly welcome!

If you want to contribute your new models, you can follow the steps below.

1. Create a folder for your model
2. The folder contains following items(you can refer to this example).
   • requirements.txt: required dependencies.
   • README.md: a brief introduction to your models
   • workflow_config_<model name>_<dataset>.yaml: a configuration which can read by qrun. You are encouraged to run your model in all datasets.
3. You can integrate your model as a module in this folder.
4. Please update your results in the above Benchmark Tables, e.g. Alpha360, Alpha158(the values of each metric are the mean and std calculated based on 20 Runs with different random seeds. You can accomplish the above operations through the automated script provided by Qlib, and get the final result in the .md file. if you don't have enough computational resource, you can ask for help in the PR).
5. Update the info in the index page in the news list and model list.

Finally, you can send PR for review. (here is an example)

Q: What's the difference between models with name *.py and *_ts.py?

A: Models with name *_ts.py are designed for TSDatasetH (TSDatasetH will create time-series automatically from tabular data). Models with name *.py are designed for DatasetH (DatasetH is usually used in tabular data. But users still can apply time-series models on tabular datasets if the columns has time-series relationships).