In recent years, domain-specific knowledge bases (KBs) have attracted more attention in academics and industry because of their expertise and in-depth representation in a specific domain. However, when constructing a domain-specific KB, one needs to address not only the challenges in constructing a general KB, but also the difficulties raised by the nature of domain-specific raw data. Considering the usability of financial Knowledge Bases (KBs) in many downstream applications, such as financial risk analysis and fraud detection, we propose a Platform for Temporal Financial KB Construction, K-Quant Platform.
K-Quant system have three main components: Knowledge Module, Quantitative Investment Module, XAI Module.
The main features of K-Quant:
The dynamic financial KB construction module consists of three main steps:
-
Evolved Knowledge Extraction: it extracts financial knowledge with evolved information from several reliable data sources.
-
Temporal Record Linkage and Conflict Resolution: it removes duplicated and conflicted knowledge based on temporal information to increase the confidence of the delivered knowledge.
-
Dynamic Knowledge Update: it proposes a temporal pattern-based inference rule learning module to maintain our constructed KB consistent with the dynamically changing world.
To handle temporal relations between companies and utilize the knowledge in quantitative investment, the K-Quant have three main functions in Quantitative Investment Framework:
-
Knowledge based stock recommendation models
We build knowledge based deep learning models for stock forecasting tasks. By extracting knowledge formed in KB construction module as relational matrix, we use them as an alternative data. For detailed models, please refer to readme of stock prediction
-
Stock recommendation ensemble models
To boost the performance of stock recommendation, we employ ensemble methods to sample the output of different deep learning models. To better capture the rapid change of stock prices, we design several dynamic ensemble methods including Online ensemble method based on resampling distribution estimation and dynamically heterogeneous ensemble methods. Both of these two methods apply for patents.
-
Incremental learning methods for stock models
The complex stock market make deep learning model invalid soon because of the distribution shift. We propose incremental learning methods to mitigate the influence of such shift.
To response the limitation of deep learning based methods without explanation of the prediction results, the K-Quant provides an explainers module for knowledge-based prediction models as well as an assessment module for assessing different models performance.
-
Interpretability Methods
We build explainers models to identify the most relevant stocks from stocks relation graph of each prediction results coming from the selected stocks. With the help of knowledge graph, sufficient information is not only used to support quantitative investment, but also leveraged to explain the results of prediction models.
-
Assessment Modules
This system consists of two parts: (1)Multi-dimensional performance evaluation for different combinations of forecasting and explanatory models.(2)Performance assessment of different investment portfolios. In the first part, users can select different forecasting and explanatory models as well as examination periods. The evaluation system will output comprehensive performance scores for different combinations.Additionally, the forecasting models in the first part will also provide recommendations of three stocks for users to consider. In the second part, users can input different investment portfolios based on their own stock selection results, and test the performance of these portfolios over different time periods.
For Quantitative Investment Framework, you need install Microsoft Qlib and download up-to-date stock time series data.
We recommend install Qlib from source and download the crowd source stock data
# install Qlib from source
pip install --upgrade cython
git clone https://github.com/microsoft/qlib.git && cd qlib
python setup.py install
# Download the stock features of Alpha360 from Qlib
# the target_dir is the same as provider_url in utils/dataloader.py
python scripts/get_data.py qlib_data --target_dir ~/.qlib/qlib_data/cn_data --region cn --version v2
For detailed usage of quantitative forecasting model, please refer to Stock prediction module
Please refer to our HiDy Dataset.
HiDy is a hierarchical, dynamic, robust, diverse, and large-scale financial benchmark KB that aims to provide various valuable financial knowledge as critical benchmarking data for fair model testing in different financial tasks. Specifically, HiDy currently contains 34 relation types and 17 entity types. The scale of HiDy is steadily growing due to its continuous updates. To make HiDy easily accessible and retrieved, HiDy is organized in a well-formed financial hierarchy with four branches, Macro, Meso, Micro, and Others.
Each instance in the dataset is in the form of quadruple (entity 1, relation , entity 2, timestamp).
If you find this system helpful in your research, please consider citing our papers that are listed below:
Cost-aware Outdated Facts Correction in the Knowledge Bases. Hao Xin, Yanyan Shen, Lei Chen. DASFAA 2024. [Paper Coming Soon]
PKBC: A Product-Specific Knowledge base Taxonomy Framework. Hao Xin, Yanyan Shen, Lei Chen. DASFAA 2024. [Paper Coming Soon]
Triple-d: Denoising Distant Supervision for High-quality Data Creation. Xinyi Zhu, Yongqi Zhang, Lei Chen, Kai Chen. ICDE 2024. [Paper Coming Soon]
T-FinKB: A Platform of Temporal Financial Knowledge Base Construction. Xinyi Zhu, Liping Wang, Hao Xin, Xiaohan Wang, Zhifeng Jia, Jiyao Wang, Chunming Ma, Yuxiang Zeng. ICDE 2023. [Paper Coming Soon]
Liping Wang, Yanyan Shen, and Lei Chen. 2023. TE-DyGE: Temporal Evolution-Enhanced Dynamic Graph Embedding Network. In Database Systems for Advanced Applications: 28th International Conference, DASFAA 2023, Tianjin, China, April 17–20, 2023, Proceedings, Part III. Springer-Verlag, Berlin, Heidelberg, 183–198. https://doi.org/10.1007/978-3-031-30675-4_13
Xinyi Zhu, Hao Xin, Yanyan Shen, and Lei Chen. HIT - An Effective Approach to Build a Dynamic Financial Knowledge Base. In Database Systems for Advanced Applications: 28th International Confer- ence, DASFAA 2023, Tianjin, China, April 17–20, 2023, Proceedings, Part II, pages 716–731. Springer, 2023. https://link.springer.com/chapter/10.1007/978-3-031-30672-3_48
Liping Wang, Jiawei Li, Lifan Zhao, Zhizhuo Kou, Xiaohan Wang, Xinyi Zhu, Hao Wang, Yanyan Shen, Lei Chen. Methods for Acquiring and Incorporating Knowledge into Stock Price Prediction: A Survey. https://arxiv.org/abs/2308.04947