SDTP: Series decomposition Transformer with period-correlation for stock market index prediction

Stock price forecasting has been always a difficult and crucial undertaking in the field of finance. Enlighted by the classic time series analysis and stochastic process theory, we propose the SDTP as a safe and reliable stock prediction model [paper]. SDTP goes beyond the baseline models and performs well in most stock markets.

1. Series decomposition layer

It is difficult to learn the complex temporal patterns directly in the financial series, and we are inspired by the work of Cleveland (1990), Taylor and Letham(2018), Wu, Xu, Wang, and Long (2021), which decomposes the series into the trend and seasonal parts.

Figure 1. Overall architecture of SDTP.

2. Period-correlation mechanism

Observing that the nodes in the corresponding positions in the period-based sub-series usually have similar properties to each other (Xu, Li, Qian, & Wang, 2022), we are inspired by Wu, Xu, et al. (2021) to use period-correlation instead of traditional self-attention, which can detect the periodic relationship between the series more reasonably.

Figure 2. Period-correlation mechanism.

• Python 3.8
• pandas==1.5.2
• torch==1.12.1
• numpy==1.23.4
• matplotlib==3.7.0
• tensorflow==2.11.0
• transformers==4.26.1

Dependencies can be installed using the following command:

pip install -r requirements.txt

Commands for training and testing the model on Dataset SH:

# SH
python -u run.py --model SDTP --data custom  --root_path ./ETT-small/    --data_path  price_SH.csv --freq d --enc_in 8 --dec_in 8 --c_out 8 --target closeIndex --train_epochs 100 --batch_size 64 --features MS --itr 3 --seq_len 5 --label_len 5 --pred_len 1 --dropout 0.05 --learning_rate 0.001 --patience 50 --moving_avg 5 --lradj type1 --e_layers 5 --d_layers 3 --d_ff 1024 --d_model 512

More parameter information please refer to run.py.

We provide a more detailed and complete command description for training and testing the model:

python -u SDTP.py --model <model> --data <data>
--root_path <root_path> --data_path <data_path> --features <features>
--target <target> --freq <freq> --checkpoints <checkpoints>
--seq_len <seq_len> --label_len <label_len> --pred_len <pred_len>
--enc_in <enc_in> --dec_in <dec_in> --c_out <c_out> --d_model <d_model>
--n_heads <n_heads> --e_layers <e_layers> --d_layers <d_layers>
--s_layers <s_layers> --d_ff <d_ff> --factor <factor> --padding <padding>
--distil --dropout <dropout> --attn <attn> --embed <embed> --activation <activation>
--output_attention --do_predict --mix --cols <cols> --itr <itr>
--num_workers <num_workers> --train_epochs <train_epochs>
--batch_size <batch_size> --patience <patience> --des <des>
--learning_rate <learning_rate> --loss <loss> --lradj <lradj>
--use_amp --inverse --use_gpu <use_gpu> --gpu <gpu> --use_multi_gpu --devices <devices>

The detailed descriptions about the arguments are as following:

The above parameters are for reference only, the actual values please refer to the code.

The required data files should be put into /ETT-small/ folder. we select three well-known stock indices, the Shanghai Composite Index (SH) from July 1, 1991 to June 30, 2020, and the Shenzhen Component Index (SZ) and the Hang Seng Index (HSI) from July 31, 1991 to May 31, 2022.
To further demonstrate the feasibility of our proposed SDTP model for the stock price prediction, we conduct the experiments with the Reservoir Computing (RC) (Wang et al., 2021) model in eight different stock indices, including DJI, the Nasdaq Composite Index, the SSE Index, the FTSE 100 Index, the Nikkei 225 Index, the NYSE Composite Index, the CAC40 index, and the S&P500 Index.

We compare our model with benchmarks in popular stock indexes, including SH, SZ, etc. Generally, SDTP achieves SOTA over previous baselines.

We will keep adding series forecasting models to expand this repo:

• TimesNet
• Informer
• Transformer
• Reformer
• CNN-BiLSTM-AM
• Reservoir Computing

The baseline code is in /baseline and /models

If you find this repo useful, please cite our paper.

@article{TAO2024121424,
title = {Series decomposition Transformer with period-correlation for stock market index prediction},
journal = {Expert Systems with Applications},
volume = {237},
pages = {121424},
year = {2024}
}

We appreciate the following github repos a lot for their valuable code base or datasets:

https://github.com/thuml/Autoformer

https://github.com/zhouhaoyi/Informer2020