Stock Trend Prediction with Multi-granularity Data: A Contrastive Learning Approach with Adaptive Fusion

0. Dependencies

   Install packages from requirements.txt.

1. Load Data using qlib

   $ cd ./load_data
   

   Download daily data:

   $ python load_dataset.py

   • Change parameter market to get data from different dataset: csi300, csi800, NASDAQ etc.

     Data Sample - SH600000 in CSI300

     

     features dimensions = 6 * 20 + 1 = 121

   Download high-frequency data:

   $ python high_freq_resample.py

   • Change parameter N to get data from different frequencies: 15min, 30min, 120min etc.

     Data Sample - SH600000 in CSI300

     

     features dimensions = 16 * 6 * 20 + 1 = 1921

2. Framework

• Pre-training Stage: Contrastive Mechanisms：./framework/models/contrastive_all_2_encoder.py
• Adaptive Multi-granularity Feature Fusion: ./framework/models/contrastive_all_2_stage.py

3. Run

$ cd ./framework

Train Pre-train model:

$ python main_contrast.py with config/contrast_all_2_encoder.json model_name=contrastive_all_2_encoder

• Add hyper-param = {values} after with or change them in config/main_model.json
• Prediction results of each model are saved as pred_{model_name}.pkl in ./out/.

Train Adaptive Multi-granularity Feature Fusion model:

$ python main_contrast_2_stage.py with config/contrast_all_2_stage.json model_name=contrastive_all_2_stage

Run Market Trading Simulation:

• Prerequisites:
  • Server with qlib
  • Prediction results

$ cd ./framework

$ python trade_sim.py

4. Records

   Records for each experiment are saved in ./framework/my_runs/.
   Each record file includes:

   config.json

   • contains the parameter settings and data path.

   cout.txt

   • contains the name of dataset, detailed model output, and experiment results.

   pred_{model_name}_{seed}.pkl

>  * contains the  `score` (model prediction) and `label`

> run.json

* contains the hash ids of every script used in the experiment. And the source code can be found in `./framework/my_runs/source/`.