Data Science / Machine Learning
Financial Forecasting & Market Trend Analysis
1976–2025 · 50 years of data
Graduate project · Sept – Dec 2025
- Python
- ARMA/SARIMA
- Diagnostics
- Holdout validation
Problem
Short-term market-trend evaluation and liquidity-related forecasting require models whose predictions can be trusted for decision support: a BlackRock equity price series for short-term trends, and a 50-year California unemployment series (1976–2025) for seasonal and cyclical signals.
Architecture
Two modeling tracks, both validated the same way. Equity prices: ARMA models with orders selected via ACF/PACF diagnostics. The macro series: SARIMA to capture seasonality and cyclical structure. Every forecast passed through a shared validation layer — holdout testing on unseen data, confidence intervals, and residual diagnostics — before being treated as usable.
Two parallel modeling tracks feed a shared validation gate. Equity price data flows into an ARMA model with ACF/PACF-based order selection. California unemployment data (1976–2025) flows into a SARIMA model. Both models' outputs pass through holdout testing, confidence-interval estimation, and residual diagnostics before producing a validated forecast.
Key Decisions
- ARMA for equity prices, with orders selected via ACF/PACF diagnostics.
- SARIMA for the 50-year macro series, to model seasonal and cyclical structure directly.
- Every model passed holdout testing, confidence-interval estimation, and residual diagnostics before its output was used for forecasting.
Results
Forecasts on both series were validated through holdout testing with confidence intervals. Residual diagnostics confirmed the models behaved as expected — residuals consistent with white noise — on both the equity series and the 50-year macro series.