Journal of Scientific Research and Reports

Rapeseed–mustard is an important oilseed crop in Bihar, and reliable yield forecasts are useful for production planning and policy support. This study developed and compared statistical, machine-learning and hybrid forecasting models for rapeseed–mustard yield in Bihar using annual historical yield data from 1950 to 2023 obtained from IndiaStatAgri. Outliers were examined using the interquartile range approach, and the series was divided into a training period (1950–2016) and a testing period (2017–2023). Stationarity was assessed using the KPSS test, and first differencing was applied before fitting ARIMA models. Candidate ARIMA specifications were evaluated using AIC and BIC, while forecast accuracy was assessed using RMSE and MAPE for the testing data. The ARIMA(0,1,1) model showed the lowest AIC among the candidate ARIMA models; however, residual diagnostics indicated significant autocorrelation, suggesting that a linear model alone did not fully capture the structure of the yield series. Therefore, hybrid models combining ARIMA with ANN and SVR were evaluated alongside standalone SVR and ARIMA models. Among the competing models, the Hybrid ARIMA–SVR model produced the best testing performance, with an RMSE of 63.52 and a MAPE of 4.50. The selected Hybrid ARIMA–SVR model forecast yields of 1257.64, 1275.00, 1263.27 and 1265.40 kg/ha for 2024, 2025, 2026 and 2027, respectively. The results indicate that the Hybrid ARIMA–SVR model can provide comparatively accurate short-term forecasts for rapeseed–mustard yield in Bihar based on historical yield patterns.