StockFormer
Hybrid trading machine integrating predictive coding and RL.
| Date | Investors | Amount | Round |
|---|---|---|---|
| N/A | €0.0 | round | |
| investor | €0.0 | round | |
* | JPY130m | Debt | |
| Total Funding | 000k | ||
StockFormer is a financial technology model rather than a formal company, presented in academic papers as a hybrid trading machine. It integrates the forward modeling capabilities of predictive coding with the policy flexibility of reinforcement learning (RL) agents. The model was introduced in a paper by Siyu Gao, Yunbo Wang, and Xiaokang Yang. Another distinct model also named Stockformer was introduced by Bohan Ma, Yushan Xue, Yuan Lu, and Jing Chen.
The core of the first StockFormer is to address a key challenge in RL-for-finance: making trading decisions that are aware of the dynamic relationships between multiple assets and their future trends. It consists of a predictive coding module with three modified Transformer branches that extract latent states of long-term and short-term future dynamics, as well as asset relations. An RL agent then adaptively fuses these states to execute an actor-critic algorithm. This entire system is jointly trained, allowing it to outperform existing methods on public financial datasets in portfolio returns and Sharpe ratios.
A separate model, also named Stockformer, is described as a price-volume factor stock selection model designed for the complex and fluctuating Chinese stock market. This version uses wavelet transformation to break down stock returns into high and low frequencies to capture both long-term trends and short-term volatility. It incorporates a Dual-Frequency Spatiotemporal Encoder and graph embedding to analyze temporal and spatial relationships between stocks. This model, whose code is open-sourced on GitHub, has demonstrated strong stability and performance in backtesting across various market conditions. There is also a mention of a Tokyo-based proptech company named StockFormer, founded in 2019, described as a real estate investment information platform, but detailed information is scarce.
Keywords: quantitative trading, reinforcement learning, predictive coding, financial modeling, stock selection model, algorithmic trading, actor-critic algorithm, transformer networks, time series analysis, portfolio optimization, financial machine learning, wavelet transform, spatiotemporal encoder, market dynamics, asset relations