Oil Market Manipulation Analysis
Overview
A modular analytical software testbed running 5 concurrent evaluation engines — SARIMAX time-series forecasting, XGBoost regression, Granger causality testing, PPO reinforcement learning, and unsupervised anomaly detection — over multi-asset market and subsidy data streams.
Approach
- Designed a modular data engine processing synthetic market indicators timed to historical macroeconomic shock points (OPEC cuts, sanctions, demand shifts).
- Engineered Granger causality testing to differentiate statistical lead-lag relationship directions from simple correlation spikes.
- Built a comparative execution framework evaluating supervised classification, unsupervised clustering, and PPO reinforcement learning agents against identical stream inputs.
Significant through lag 8 (p<0.01). Intervention leads hidden subsidy by up to eight months.
Results
Metrics — sourced from the repo (README / model card / test output)
- Evaluation engines
- 5 (SARIMAX, regression sweep, Granger, anomaly detection, PPO)
- Regression model families swept
- 7 (Linear, Ridge, KNN, Decision Tree, Random Forest, XGBoost, LightGBM)
- Real macro events anchoring synthetic data
- 4 (2016 OPEC cuts, 2018 sanctions, 2020 COVID demand collapse, 2022 Russia sanctions)
- Deployment status
- not yet deployed
next step per repo: validate against a labeled or real-world dataset
Achieved multi-model alignment across independent algorithmic lenses while maintaining strict validation boundaries before real-world dataset deployment. The dataset is explicitly synthetic — documented as such in the repo rather than presented as observed market data — so the metrics below describe engineering scope, not a deployed-accuracy claim.