I already have a working PyTorch model that forecasts petrol-price time-series data, yet its accuracy is not where it needs to be. The core goal is to push prediction performance higher without bloating runtime or resource usage. You will receive: • the current repo (Python 3, PyTorch 2.x) I would like you to: • craft smarter feature engineering steps (calendar effects, rolling stats, lags, etc.) • tune hyperparameters systematically (learning rate, hidden sizes, sequence length, regularisation) • rethink or refine the model architecture if it unlocks accuracy gains (e.g., temporal CNN layers, attention, Transformer-style blocks) • keep RAM and CPU footprints reasonable so the model can still train on a mid-range workstation Please document every change clearly inside the code and in a short README so I understand both the rationale and how to reproduce your results.