PyTorch Crypto Price Prediction Model
Budget: $750 – $1,500 USD
I have a large set of historical price data for several major cryptocurrencies and want a deep-learning model, built in PyTorch and trained on GPU, that can reliably forecast the next price move. The core objective is straightforward: predict future prices based solely on past market behavior; no fundamental or news inputs will be supplied for this first phase.
What I will hand over
• Cleaned CSV files containing timestamped open, high, low, close, volume and others for bitcoin
What I need back
1. Well-structured PyTorch code (Python 3.10+) that loads the data, constructs the model, trains, validates, and saves the best weights.
2. A short README explaining the data pipeline, hyper-parameters, and how to reproduce results on any CUDA-enabled machine.
3. Evaluation notebook or script that reports at least MAE, RMSE, and directional accuracy on an unseen test split.
4. Sample inference routine that accepts the most recent window of prices and returns the next predicted value.
Acceptance criteria
• Training completes without errors on a single GPU and converges to a loss trend that clearly improves on a naïve persistence baseline.
• Reproducible results using the provided random seed.
• Code passes a quick review for clarity, modularity, and adherence to PyTorch best practices (no deprecated APIs).
Choose the exact architecture—Transformer—as long as it is justified in the README and can be tuned later. I’m open to additional ideas such as feature engineering or ensembling once the core model is stable. Let me know your estimated timeline for an initial working version and any questions about the dataset or environment.
What I will hand over
• Cleaned CSV files containing timestamped open, high, low, close, volume and others for bitcoin
What I need back
1. Well-structured PyTorch code (Python 3.10+) that loads the data, constructs the model, trains, validates, and saves the best weights.
2. A short README explaining the data pipeline, hyper-parameters, and how to reproduce results on any CUDA-enabled machine.
3. Evaluation notebook or script that reports at least MAE, RMSE, and directional accuracy on an unseen test split.
4. Sample inference routine that accepts the most recent window of prices and returns the next predicted value.
Acceptance criteria
• Training completes without errors on a single GPU and converges to a loss trend that clearly improves on a naïve persistence baseline.
• Reproducible results using the provided random seed.
• Code passes a quick review for clarity, modularity, and adherence to PyTorch best practices (no deprecated APIs).
Choose the exact architecture—Transformer—as long as it is justified in the README and can be tuned later. I’m open to additional ideas such as feature engineering or ensembling once the core model is stable. Let me know your estimated timeline for an initial working version and any questions about the dataset or environment.