Supervised and reinforcement learning applied to forecasting — trained models with committed weights, not notebooks.
The only trained models in my public work, all three in one repository:
- scikit-learn RandomForest with a committed
model.pkl - Keras LSTM for sequence modelling
- Hand-rolled PPO actor-critic — a reinforcement learning agent written from scratch rather than pulled from a library
Honest status: the repository does not currently run — a module it imports is missing, a base class is never imported, it carries a hardcoded absolute path, and 17 files are zero bytes including all four test files. The modelling is real; the packaging is broken, and that needs fixing before anyone judges the models.
https://github.com/pranay123-stack/Smarkets_Sports_Quant_Trading
Sentence-transformer embeddings matching the same real-world event across Kalshi and
Polymarket, inside prediction-market-arbitrage-bot. Solving entity resolution with
embeddings rather than string matching is the interesting part.
https://github.com/pranay123-stack/Algorithmic-Trading-Projects
Multi-signal prediction with injury-driven features and cross-market momentum-lag detection. 18 tests. Has never traded — dry-run by default, stated openly in both the README and the code.
https://github.com/pranay123-stack/nba-prediction-edge
Road accident severity prediction — a scikit-learn MLP over a road-accident dataset, with the trained pipeline and analysis. Roughly 29 Python/notebook files.
Equity price prediction (Korean market) — a desktop prediction platform against a broker Open API, ~175 source files, delivered under a commercial engagement. The client owns it; it stays private.
There is no computer vision, no fine-tuning and no deployed inference service in my public work. What is here is classical ML and RL applied to market and outcome forecasting.
Tech: Python, scikit-learn, TensorFlow/Keras, sentence-transformers, pandas
Related: AI & LLM Engineering · Agentic Applications