Request for the trained BTC 5-minute predictor model and reproducibility package
Hi Philippo,
Thank you for publishing your BTC 5-minute Polymarket Predictor. I am contacting you on behalf of the ZEUS research project, an independent engineering effort focused on building a statistically robust paper/shadow analytics and decision-support system for Bitcoin Up or Down markets on Polymarket.
Over the course of this project, we have already built and validated most of the technical infrastructure required for serious research. The system can identify the correct recurring BTC markets and their outcome tokens, collect synchronized Bitcoin reference data and Polymarket market observations, enforce strict freshness and execution-quality checks, associate candidate decisions with final market outcomes, and produce reproducible analytical reports.
All current work is strictly paper/shadow. Order submission, signing, authenticated trading, and real-money execution remain disabled.
The technical infrastructure is now functioning reliably. Our main difficulty is no longer collecting market data or maintaining the system. The challenge is the analytical layer: identifying a predictive signal that remains stable across different Bitcoin market regimes, avoids overfitting, and retains positive expectancy after realistic spread, depth, latency, fees, and execution assumptions.
We have completed long-running forward-shadow data-collection sessions and accumulated a substantial internal research history. These tests confirmed that the technical system can operate reliably, but they also showed that the current analytical approach is sensitive to market regime. In particular, a signal that may behave reasonably during range-bound conditions can deteriorate sharply during persistent directional movement.
Your model is especially relevant to us because it was designed specifically for BTC 5-minute Polymarket markets and appears to combine short-horizon Bitcoin momentum, Polymarket microstructure, order-book information, and market-state features. This is closely aligned with the analytical direction we are currently developing.
Access to the exact trained artifact and its reproducibility details could help us evaluate a substantially stronger analytical baseline without duplicating all of your original training work. We would integrate it only into an isolated, safety-locked paper/shadow environment and compare it independently against our existing approaches.
I can see the training and inference source code, feature engineering, sample data, and reported metrics in the Hugging Face repository. However, I could not find the trained model artifact required by predict.py, such as models/lgbm_swing_v8.pkl or the latest model version currently recommended.
Would you be willing to provide or upload the following?
The latest trained model artifact used for the published results, preferably v8 or whichever version you consider the current reference model.
The exact SHA-256 hash of that model file.
The exact source revision used to train it, ideally a Git commit SHA, release tag, or source archive with its own SHA-256.
The exact ordered feature list expected by the trained model.
Any preprocessing, calibration, encoders, scalers, or auxiliary objects required for inference.
The Python and dependency versions used for training and inference, especially LightGBM, pandas, NumPy, scikit-learn, and joblib.
The decision threshold associated with the reported precision, recall, and expected-value results.
Any notes about the chronological training and validation split, excluded markets, label construction, or known limitations necessary to reproduce the published evaluation.
Confirmation that the trained artifact and source may be used under the repository’s stated license for non-custodial paper/shadow research.
If the complete training dataset cannot be shared, that is entirely understandable. The trained model artifact, exact source revision, ordered feature contract, dependency versions, and calibration details would already be extremely valuable.
We would evaluate the model independently and conservatively. The first phase would be offline historical validation, followed by a zero-order forward-shadow test. We would explicitly test for temporal leakage, regime dependence, calibration drift, spread, available depth, latency, fees, and out-of-sample stability.
We would not use the model for real-money trading unless it first demonstrated robust performance across multiple independent historical and forward-shadow tests.
We would also be happy to share a concise summary of our independent validation findings with you, including where the model appears to generalize well and where market-regime or execution constraints reduce its performance.
Thank you again for making the project public. A reproducible release of the trained model would be extremely helpful to our research.
Best regards,
ZEUS Research Project
Independent Polymarket Paper/Shadow Research