Built
SARIMA Price Forecasting
Predicts a monthly home price index 12 months ahead with 1.17% average error.
The context
A forecasting study using 176 months of Case-Shiller home price data, from 2012 to 2026.
What I built
- Checked stationarity with the ADF test, then applied regular and seasonal differencing
- Compared 36 model combinations and chose SARIMA(2,1,0)(0,1,0,12) by AIC
- Confirmed with the Ljung-Box test that no clear pattern was left in the errors
- Tested on 8 rolling 12-month periods: 1.17% MAPE, against 1.57% for ARIMA and 1.37% for the best simple baseline
- Also tested ARIMA on RELIANCE stock prices, where it did no better than a naive forecast
The result
Beat ARIMA and the best simple baseline, and showed clearly where this kind of model stops working.