MgRL-CE: A study of time series prediction based on Multi-granularity Residual Learning and Confidence Estimation.
This repository provides the code for my paper "A study of time series prediction based on Multi-granularity Residual Learning and Confidence Estimation."
MgRL-CE/
├── images # All used images of this repo.
├── MgRL_Framework.png # The MgRL FrameWork (without CE).
├── CE.png # The Confidence Estimation Module.
├── Elect_Data_Distribution.png # The distribution of uci electricity dataset.
└── MgRL_CE_Images.pptx # The raw images of MgRL_CE.
├── datasets
├── datasets_preprocess
├── elect_preprocess.py # The preprocess code of UCI electricity dataset (download from Web).
├── lob_preprocess # The preprocess package of Future LOB dataset (download from Qlib).
├── price_alignment_features # The paf algorithm.
└── lob_preprocess.py # The preprocess code of Future LOB dataset (download using Akshare).
└── index_preprocess.py # The preprocess code of CSI300 index dataset.
├── elect_dataset.py # The torch.Dataset of UCI electricity dataset (after preprocessing).
├── lob_dataset.py # The torch.Dataset of Future LOB dataset (after preprocessing).
└── index_dataset.py # The torch.Dataset of CSI300 index dataset (after preprocessing).
├── configs # The train&prediction config files of 3 datasets.
├── elect_config.py # Config file of UCI electricity dataset.
├── lob_config.py # Config file of UCI electricity dataset.
└── index_config.py # Config file of CSI300 index dataset.
├── models # The MgRL-CE models and Comparison Methods.
├── MgRL.py # The Multi-Granularity Residual Learning Net.
├── comparison_methods # All comparison methods.
├── gru.py # The Comparison Methods 1 & 9 & 10 & 11: GRU Class
├── lstm.py # The Comparison Methods 2: LSTM.
├── transformer.py # The Comparison Methods 3: Transformer.
├── deepar.py # The Comparison Methods 4: DeepAR (Will be Updated 🔥 !).
├── informer.py # The Comparison Methods 5: Informer (Updating 🔥 !).
├── sfm.py # The Comparison Methods 6: SFM.
├── alstm.py # The Comparison Methods 7: ALSTM (Will be Updated 🔥 !).
└── adv_alstm.py # The Comparison Methods 8: ADV-ALSTM (Will be Updated 🔥 !).
├── ablation_methods # All ablation methods.
├── mg_add.py # The Ablation Method 1: Mg_Add.
├── mg_cat.py # The Ablation Method 2: Mg_Cat.
└── modules.py # The modules of ablation models.
├── loss.py # The loss function of MgRLNet and MgRL_CE_Net.
├── metric.py # The metrics of y_ture and y_pred.
└── modules.py # The modules of model.
├── train_pred_MgRL.py # Training and Prediction code of `MgRLNet` and `MgRL_CE_Net` for 3 datasets.
├── train_pred_CM.py # Training and Prediction code of Comparison Methods for 3 datasets.
└── utils.py # Some util functions.
This study extensively performs experiments on 3 Real-World Datasets to verify the feasibility of the proposed MgRL-CE. You can DOWNLOAD the raw datasets from the following links, and here are also some description about the datasets.
-
UCI electricity dataset (ELECT). Could be downloaded from HERE !
The UCI electricity dataset collects electricity consumption of each 15 minutes (unit: kW*15min) from a total of 370 clients over a 4-year period from 2011 to 2014, some of which were created after 2011, and all missing data on electricity consumption for these clients are filled with ZEROS !
-
Future Limit Order Book dataset (LOB). Could be downloaded from the public
Qlibplatform, HERE !The Future Limit Order Book dataset collects high-frequency trading data of CSI 300 stock index future (IF_M0), including Limit Order Book (LOB) with 5 levels for both ask and bid direction. The trading frequency is 0.5 seconds. The dataset range from Jan. 4, 2022 to Dec. 30, 2022, covering all 242 trading days in 2022, with 28,800 trading records for each trading day.
-
CSI300 index dataset (INDEX). Could be downloaded based on the public
AKSharetoolkit, HRER !The CSI300 index dataset collects 1 minute high-frequency trading data for the CSI 300 stock index (000300.SH) from publicly available data sources. The collection interval is from the beginning of 2016 to the end of 2023, containing a total of 1,945 trading days over an 8-year period, with 240 trading records for each trading day.
After downloading the datasets following the Dataset Acquisition, data preprocessing is needed to get the structured dataset. I have released preprocess code for datasets, please read them carefully and follow the guidelines in the top comment rather than running the shell command directly ! I have also released torch.Dataset code for datasets.
-
UCI electricity dataset (ELECT).
In order to minimize the interference caused by missing data, this study intercepts the sample data from the original dataset for the 3-year period from 2012 to 2014, and excludes the clients with more than 1 day of missing data in the interval, and finally retains the electricity consumption data of 320 clients. The target task of this paper is to predict the next day's electricity consumption of each client, and the dataset is divided into training set, validation set and test set according to the time sequence, which covers 24 months, 6 months and 6 months, respectively. The feature data input to the network has 5 kind of granularity: 1 day (coarsest), 12 hours, 4 hours, 1 hour and 15 minutes (finest).
ATTENTION: During the preprocessing I have also changed the unit of data from kW*15min to kWh and adjusted the scale of data distribution by dividing each client data by their daily electricity consumption on the first day.
-
Future Limit Order Book dataset (LOB).
Similarly, the LOB dataset is divided in chronological order: the training, validation, and test sets cover 8, 2, and 2 months, respectively. In this study, the original LOB data is modeled directly, i.e., only the 20 basic features of price and volume from 1 to 5 ticks in both ask and bid directions are used, and no other factors are constructed manually. The objective is to predict the minute frequency return of future, i.e.,
$y=log(MidPrice_{T+1}/MidPrice_{T})*10^{4}$ , where$MidPrice_{t} = (Price_t^{ask} + Price_t^{bid}) / 2$ denotes the average of the 1 level ask price and bid price in the minute$t$ . There are 5 types of input feature granularity: 1 minute (coarsest), 30 seconds, 10 seconds, 1 second and 0.5 seconds (finest). All feature data were normalized by the Z-Score method. -
CSI300 index dataset (INDEX).
The training, validation, and test sets span 6 years (2016 to 2021), 1 year (2022), and 1 year (2023), respectively, in chronological order. Six commonly used market factors are extracted as feature inputs, including high price, opening price, low price, closing price, volume and turnover, and all features are normalized by Z-Score method before inputting into the model. This paper also chooses the daily return of the stock as the prediction target, i.e.,
$y=(P_{T+2}/P_{T+1}-1)\times100$ where$P_t$ stands for the average price of the CSI 300 stock index on the$t$ -th day. This dataset also has 5 feature granularities: 1 day (coarsest), 1 hour, 15 minutes, 5 minutes, and 1 minute (finest).
There are some differences between the different datasets during Training and Prediction. Please carefully set the config files of different datasets following my example.
-
UCI electricity dataset (ELECT).
-
Future Limit Order Book dataset (LOB).
-
CSI300 index dataset (INDEX).
This study compares the proposed method with numerous other methods. The competitive baselines i compared can be categorized into 4 Groups:
GROUP 1. General Time Series Forecasting Models (using single granularity)
- GRU, HERE. Kyunghyun Cho, et al. 2014, Ref. Code.
- LSTM, HERE. Sepp Hochreiter, et al. Neural computation 1997, Ref. Code.
- Transformer, HERE. Ashish Vaswani, et al. NeurIPS 2017, Ref. Code.
- DeepAR, HERE (Will be Updated 🔥). Salinas D, et al. Int. J. Forecasting 2020, Ref. Code, Ref. Video.
- Informer, HERE (Updating 🔥). Zhou H, et al. AAAI 2021, Ref. Code, Ref. Video
GROUP 2. Current TOP Models for Stock Trend Prediction (using single granularity)
- SFM, HERE. Liheng Zhang, et al. KDD 2017, Ref. Code.
- ALSTM, HERE (Will be Updated 🔥). Yao Qin, et al. IJCAI 2017, Ref. Code.
- ADV-ALSTM, HERE (Will be Updated 🔥). Feng F, et al. IJCAI 2019, Ref. Code.
GROUP 3. Model Variants (using different granularity of data)
- Fine-Grained GRU, HERE. using only finest-grained data.
- Multi-Grained GRU, HERE. using the concatenation of two granularity data.
- Ensemble, HERE (Updating 🔥). ensemble result for five independent training models with different granularity data.
GROUP 4. Two Ablation Models for MgRL_CE (using different granularity of data)
- MgRL, HERE, not containing the confidence estimation mechanism in MgRL_CE.
- MgRL_Attention, HERE, replacing the confidence estimation mechanism in MgRL_CE with the classical soft attention mechanism.
To facilitate the comparison of all Comparison Methods, i built the training and prediction frameworks for each comparison methods of 3 datasets !
python3 train_pred_CM.py --dataset dataset_name --method method_name