Research map / Time series / Long-term forecasting
StockMixer: A Simple Yet Strong MLP-Based Architecture for Stock Price Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes StockMixer, a lightweight MLP-based model for stock return forecasting that mixes indicators, multi-scale time patches (with causal masking), and stocks via a market-bottleneck, outperforming RNN, GNN, and hypergraph baselines on NASDAQ, NYSE and S&P500.
Problem
Stock price forecasting on multivariate stock data, where complex hybrid neural architectures (RNN/GNN/Transformer) overfit or are hard to optimize given limited stock data (~250 trading days per year).
Approach
Replace standard MLP-Mixer stock mixing with a stock-to-market-to-stock bottleneck (learnable hyperedge-like compression to m market states), and use multi-scale average-pooled time patches with causal (upper-triangular) time mixing, plus indicator mixing before time mixing.
Key results
- On NASDAQ, StockMixer reaches IC 0.043 and RIC 0.501 versus 0.039/0.451 for the best baseline STHAN-SR.
- On S&P500, StockMixer reports IC 0.041 and Sharpe Ratio 1.586, the best among compared methods (ESTIMATE SR 1.547).
- Average relative gains of 7.6%, 10.8% and 10.9% in two rank metrics and Sharpe ratio are reported.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add indicator mixing (ablation) | StockMixer · NASDAQ | IC: 0.040 -> 0.043 | Helped |
| Add indicator mixing (ablation) | StockMixer · NYSE | IC: 0.027 -> 0.029 | Helped |
| Add time mixing (ablation) | StockMixer · NASDAQ | IC: 0.018 -> 0.043 | Helped |
| Add time mixing (ablation) | StockMixer · NYSE | IC: 0.016 -> 0.029 | Helped |
| Add stock mixing (ablation) | StockMixer · NASDAQ | IC: 0.037 -> 0.043 | Helped |
| Add stock mixing (ablation) | StockMixer · NYSE | IC: 0.026 -> 0.029 | Helped |
| Replace the indicator-and-time encoder with LSTM while retaining stock mixing (ablation) | StockMixer · NASDAQ | IC: 0.041 -> 0.043 | Hurt |
| Replace the indicator-and-time encoder with LSTM while retaining stock mixing (ablation) | StockMixer · NYSE | IC: 0.030 -> 0.029 | Mixed |
| Use ReLU or HardSwish instead of GELU (ablation) | StockMixer mixing blocks · NASDAQ | model performance | Helped |
| Change the activation to Sigmoid or tanh (ablation) | StockMixer mixing blocks · NASDAQ | model performance | Hurt |
| Add Layer Normalization (ablation) | StockMixer mixing blocks · stock datasets | model performance | No clear effect |
| Vary the lookback window length (ablation) | StockMixer · NASDAQ, NYSE, S&P500 | IC | Mixed |
Surprising findings
- Standard MLP-Mixer and vanilla linear/MLP time-series models perform worse than basic RNN models on stock data.
- GELU, the default in vision MLPs, underperforms ReLU and HardSwish on stock sequences.
Limitations
- Hyperparameter m must be tuned per market via grid search and is sensitive (e.g., S&P500 degrades when m exceeds 10).
- Slight performance degradation on NYSE, the largest stock pool, suggesting insufficient inductive bias for larger candidate sets.
Experimental setup
Datasets: NASDAQ NYSE S&P500
Benchmarks: NASDAQ NYSE S&P500
Metrics: IC RIC Precision@N Sharpe Ratio
Baselines: LSTM ALSTM RGCN GAT RSR-I STHAN-SR ESTIMATE Linear
Builds on: MLP-Mixer RSR STHAN-SR ESTIMATE ALSTM
Closest papers by meaning
- MASTER: Market-Guided Stock Transformer for Stock Price Forecasting (AAAI 2024 · similarity 0.88)
- HyperIV: Real-time Implied Volatility Smoothing (ICML 2025 · similarity 0.85)
- DHMoE: Diffusion Generated Hierarchical Multi-Granular Expertise for Stock Prediction (AAAI 2025 · similarity 0.85)
- CI-STHPAN: Pre-trained Attention Network for Stock Selection with Channel-Independent Spatio-Temporal Hypergraph (AAAI 2024 · similarity 0.84)
- TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (ICLR 2024 · similarity 0.84)
- Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.