RuneCast

RuneCast is a deep learning price forecasting system for the RuneScape Grand Exchange, built with PyTorch and trained on multivariate time series data scraped from live game APIs.

The core insight driving the project is that crafted items in RuneScape often have layered supply chains; finished goods depend on intermediate products and raw materials. This project aims to predict the price of glorious bars, items that are used to craft an end-game armour set, which depend on many lower level ores and stone spirits, each with their own price volatility.

Rather than treating each item in isolation, RuneCast models these dependencies explicitly through a configurable precursor depth system, feeding the full upstream price history into an LSTM at training and inference time.

The pipeline covers the full ML lifecycle: automated data collection via the weirdgloop Exchange API, sequence construction and standard scaling, LSTM training with early stopping and model checkpointing, and hyperparameter optimisation using Optuna. Predictions are generated against live-scraped price data, with results reported as both a price forecast and directional signal.

The project was originally my MSc dissertation in Intelligent Systems and Robotics, rebuilt from scratch with a cleaner architecture, corrected inference logic, and a more rigorous evaluation framework. It demonstrates applied time series forecasting, supply chain dependency modelling, and end-to-end ML pipeline design in a domain with genuine market complexity.

Due to potential risks to the in-game economy and the players of RuneScape, I am choosing not to publicly release the code for this project. If you're interested, I may share it privately at my own discretion.