Dynamic group fusion transformer for financial time series prediction: An ablation study
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Forecasting financial time series is particularly challenging since market data is complicated and non-stationary, and it is necessary to identify both short-term momentum and long-term structural patterns. This work develops the Dynamic Group Fusion Enhanced Transformer (DGFET), a new approach that integrates adaptive feature group fusion and selective information processing. The suggested DGFET architecture has Group-FiLM adapters that use Dynamic Group Fusion techniques for adaptive feature transformation to manage market, fundamental, technical, and sentiment feature groups. We assess the model using four unique labeling strategies: short-horizon momentum (binary/ternary) and triple-barrier (binary/ternary), which represent various temporal horizons and forecasting methods. Our ablation analysis, conducted on a comprehensive EUR/USD dataset from 2010 to 2023 with 88 features, demonstrates that the proposed method consistently outperforms baseline LSTM and standard transformer models across all prediction objectives. The improved architecture has a higher overall performance, with an F1-macro score of 0.5356 and a ROC-AUC of 0.6612. It also works very well for short-horizon momentum binary classification (F1: 0.7219, ROC-AUC: 0.8105). The results show that adaptive feature fusion works better than traditional designs when combined with dynamic group selection. The best configurations depend on the specific prediction job. Our results underscore the imperative of task-specific architectural design in financial machine learning applications, especially for methodologies necessitating varied temporal horizons and prediction granularities.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.021 | 0.014 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.002 | 0.006 |
| Science and technology studies | 0.002 | 0.001 |
| Scholarly communication | 0.001 | 0.003 |
| Open science | 0.002 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it