Kronos BTC Directional Forecasting Models
Public checkpoints from an interval-aware adaptation study of the Kronos financial foundation model for BTCUSDT directional forecasting.
This model repository accompanies a Spring 2026 Senior Graduation Project II at the College of Information Technology, United Arab Emirates University.
Author: Abdelrahman Osman Supervisor: Dr. Shengcai Liao
Project links
- Source code and complete evaluation artifacts
- Original editable PowerPoint poster
- Research poster preview
- Reproducibility guide
- Experimental report
Selected results
| Horizon | Selected adaptation | Window | N | Accuracy | Baseline | Margin | Frictionless 1 BTC PnL |
|---|---|---|---|---|---|---|---|
| 5 minutes | Native direction head | 2026-01-01 to 2026-03-09 | 19,554 | 51.40% | 50.16% | +1.24 pp | +5,836.58 USDT |
| 1 hour | 50% pruned direction head | 2026-01-01 to 2026-03-04 | 1,498 | 53.67% | 50.73% | +2.94 pp | +6,069.51 USDT |
| 1 day | Last-6-layer consecutive-path adapter | 2026-01-01 to 2026-03-16 | 75 | 62.67% | 53.33% | +9.33 pp | +28,035.71 USDT |
These are historical research observations on different windows, not direct like-for-like horizon comparisons. The PnL calculation assumes frictionless execution with a fixed 1 BTC position and excludes fees, spread, slippage, latency, funding, liquidity limits, and market impact.
The 62.67% daily figure is the best observed 75-row snapshot. A later extension to 88 rows through 2026-03-29 reached 57.95% against a 56.82% majority baseline.
Repository contents
The repository publishes every saved project weight from the supplied archive: 11 files, 3,338,377,256 bytes in total, representing 10 unique binary objects. One final 1-minute rollout checkpoint is byte-identical to its best checkpoint; both public paths are retained for archive completeness.
Selected horizon models
| Model | Path | Format |
|---|---|---|
| 5-minute native direction head | models/5min-native-direction-head/best_direction_model.pt | PyTorch checkpoint |
| 1-hour 50% pruned direction head | models/1h-pruned50-direction-head/best_direction_model.pt | PyTorch checkpoint |
| 1-day last-6 path adapter | models/1d-last6-path-adapter/model.safetensors | Safetensors |
Required supporting models
| Model | Path |
|---|---|
| Fine-tuned 5-minute Kronos predictor | models/5min-kronos-base/ |
| Fine-tuned 5-minute tokenizer | models/5min-tokenizer/ |
| Upstream daily tokenizer dependency | NeoQuasar/Kronos-Tokenizer-base |
Additional experimental models
models/1m-trainall-2025h1/models/1m-retrain-2017-2024/
weights_manifest.json records the source archive path, public path, byte size,
role, and SHA-256 of every weight. checksums.sha256 supports direct integrity
verification.
Download
from huggingface_hub import snapshot_download
model_root = snapshot_download(
repo_id="abdelrahman964/kronos-finetuning",
allow_patterns=["models/**", "weights_manifest.json", "checksums.sha256"],
)
print(model_root)
Loading notes
The safetensors predictor directories include their architecture configs. The
direction-head .pt files use the custom implementation in
finetune_csv/kronos_direction_model.py
and the evaluator in
finetune_csv/eval_direction_model.py.
Some direction-head metadata retains stale training-machine paths. Use the
evaluator's explicit --predictor-path and --tokenizer-path arguments with
the public supporting directories rather than relying on embedded paths.
PyTorch .pt files use pickle-based serialization. Load them only from this
verified repository and check their SHA-256 values before use.
Intended use
- Reproducing the saved BTCUSDT direction experiments
- Comparing horizon-specific adaptation methods
- Studying moderate structured pruning as regularization
- Academic review of model artifacts and evaluation evidence
- Research prototypes and paper-trading experiments
Limitations
- The checkpoints were evaluated on one trading pair and limited market windows.
- Repeated model selection can inflate the best observed result.
- Results may degrade under distribution shift and realistic execution costs.
- The checkpoints do not provide investment advice or guaranteed profitability.
- Live exchange execution requires independent risk controls and user-supplied credentials.
License and attribution
The repository is released under the MIT License and retains attribution to the upstream Kronos project and paper:
Yu Shi et al. Kronos: A Foundation Model for the Language of Financial Markets. arXiv:2508.02739.
Users should also review the license and terms of the upstream Kronos models and their market-data provider.