Research map / Time series / Foundation models for time series
Transfer Learning in Financial Time Series with Gramian Angular Field (Student Abstract)
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.
Applies Gramian Angular Field transformations to time series before computing domain similarity functions, to improve source-domain selection for transfer learning of DNN and LSTM models on Hong Kong stock data.
Problem
Transfer learning in financial time series can suffer negative transfer when source domains are poorly chosen; existing similarity functions may not align domains well.
Approach
Transform 1-D financial time series into GAF images before computing standard domain-discrepancy similarity functions (e.g., Coral, CMD, MMD), so source selection uses angular-temporal representations; Coral (GAF) for DNN and CMD (GAF) for LSTM reduce prediction error the most.
Key results
- Coral (GAF) gives the lowest prediction error for DNN in single-source transfer to the bank domain, and CMD (GAF) does so for LSTM.
- CMD consistently minimized prediction errors and performed well in multi-source transfer learning.
- GAF benefits were less pronounced for DNNs, and GAF effectiveness varied in multi-source settings with mixed industries.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Coral (GAF) for source-domain selection | DNN transfer learning · Hong Kong stock market, single-source Bank to Bank | prediction error | Helped |
| Use CMD (GAF) for source-domain selection | LSTM transfer learning · Hong Kong stock market, single-source Bank to Bank | prediction error | Helped |
| Use GAF-based similarity functions, including Coral (GAF) and CMD (GAF) | DNN and LSTM transfer learning · Hong Kong stock market, single-source Other Industries to Bank | prediction error | Helped |
| Use CMD for source-domain selection | DNN and LSTM transfer learning · Hong Kong stock market, multi-source transfer learning | prediction error | Helped |
| Use ARE (GAF) for source-domain selection | LSTM transfer learning · Hong Kong stock market, multi-source Other Industries to Bank | prediction error | Helped |
| Use GAF-based similarity functions | DNN and LSTM transfer learning · Hong Kong stock market, multi-source transfer learning with mixed-industry sources | prediction error | Mixed |
Surprising findings
- GAF-based similarity functions were less effective in multi-source transfer when sources included mixed industries.
- The benefit of GAF transformation was less pronounced for DNN than for LSTM.
Limitations
- Computational complexity of GAF transformations is noted as a concern.
- Results depend on model and scenario; GAF benefit not uniform across settings.
- Evaluation limited to Hong Kong stock market data; generalizability untested.
Experimental setup
Datasets: Hong Kong stock market time series (bank and other industries)
Benchmarks: Hong Kong stock market time series prediction (bank target domain)
Metrics: prediction error
Baselines: Euclidean distance Coral CMD Wasserstein distance Pearson DTW TWED MMD ARE PSNR SSIM
Builds on: Gramian Angular Field (Wang and Oates 2015) Central Moment Discrepancy CORAL
Closest papers by meaning
- Towards a General Time Series Forecasting Model with Unified Representation and Adaptive Transfer (ICML 2025 · similarity 0.85)
- Conditional Diffusion Model with Nonlinear Data Transformation for Time Series Forecasting (ICML 2025 · similarity 0.85)
- From GARCH to Neural Network for Volatility Forecast (AAAI 2024 · similarity 0.84)
- Market-GAN: Adding Control to Financial Market Data Generation with Semantic Context (AAAI 2024 · similarity 0.84)
- Multi-Scale Finetuning for Encoder-based Time Series Foundation Models (NeurIPS 2025 · similarity 0.84)
- Towards Financial World Modeling (NeurIPS 2026 · similarity 0.83)
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.