ML research · Quant systems · Praedixa founder
STEVEN PVR — FORECASTING SYSTEMS
My work spans demand forecasting, quantitative finance, reproducible ML pipelines and Praedixa, an AI startup starting with demand and staffing forecasts for perishable food operations.
ARIMA-EGARCH + LightGBM
39-page research report, EN version for YC.
Praedixa Research — demand forecasting
Research repository for Praedixa's wedge: anticipating demand and operational needs in perishable food businesses.
ARIMA-EGARCH + LightGBM — S&P 500 volatility
Econometrics + ML study testing the contribution of a conditional log-sigma feature inside a multi-asset LightGBM model.
S&P 500 broker-aware forecasting
Cross-sectional forecasting and portfolio translation pipeline for XTB-tradable S&P 500 stocks.
GitHub / StevenPvr
Selected work
Each project is presented as evidence of method: problem, architecture, metrics and limitations. This is meant to show what was actually built, not inflate the narrative.
Praedixa Research — demand forecasting
Research repository for Praedixa's wedge: anticipating demand and operational needs in perishable food businesses.
ARIMA-EGARCH + LightGBM — S&P 500 volatility
Econometrics + ML study testing the contribution of a conditional log-sigma feature inside a multi-asset LightGBM model.
S&P 500 broker-aware forecasting
Cross-sectional forecasting and portfolio translation pipeline for XTB-tradable S&P 500 stocks.
Numerai Classic research stack
Research infrastructure for Numerai Classic: versioned ingestion, deterministic paths, tests and era-based validation doctrine.
Bakery sales forecasting
Product-by-product ARIMA baseline on bakery sales with temporal split, Optuna and evaluation artifacts.
HumanWeight regression benchmark
Ridge, RandomForest and LightGBM benchmark with preprocessing, Optuna, metrics and interpretability.
MentalHealth CatBoost pipeline
CatBoost pipeline on self-reported questionnaire data with logistic baseline, SHAP and equity analysis.
Flagship research
ARIMA-EGARCH + LightGBM for S&P 500 volatility forecasting.
The report tests whether a conditional log-sigma signal from ARIMA-EGARCH improves LightGBM volatility forecasts. The dedicated page covers the protocol, results, statistical tests and limitations.
Research case study
Conditional volatility as a machine-learning feature.
Current wedge
Praedixa starts with demand and staffing forecasting.
The goal is not to replace existing systems. Praedixa sits above POS, inventory, scheduling and external signals to better anticipate volumes, reduce waste and stockouts, and improve margins with measurable ROI.
Initial market: restaurant franchisees, bakeries, snacking, coffee shops and perishable food operators.
Short-term angle: demand, operational needs, material cost and staffing.
Long-term vision: an operational decision layer, without prematurely selling full automation.