Autoregressive Integrated Moving Average (ARIMA)-Based Outlier Treatment and Intervention Modelling for Forecasting Indian Standard Natural Rubber Grade 20 (ISNR 20) Prices
K. Shameem Fathima *
Department of Agricultural Statistics, College of Agriculture, Vellanikkara, Kerala Agricultural University, 680656, Kerala, India.
K. C. Ayyoob
Department of Agricultural Statistics, College of Agriculture, Vellanikkara, Kerala Agricultural University, 680656, Kerala, India.
Anil Kuruvila
Department of Agricultural Economics, College of Agriculture, Vellayani, Kerala Agricultural University, 695522, Kerala, India.
Nair Sunil Appukuttan
Department of Plantation, Spices, Medicinal and Aromatic Crops, College of Agriculture, Vellanikkara, Kerala Agricultural University, 680656, Kerala, India.
Pratheesh P. Gopinath
Department of Agricultural Statistics, College of Agriculture, Vellayani, Kerala Agricultural University, 695522, Kerala, India.
*Author to whom correspondence should be addressed.
Abstract
Background: Natural rubber prices exhibit substantial fluctuations, creating challenges for accurate forecasting and market decision-making. Outliers and structural changes associated with policy interventions may influence time-series modelling and forecast accuracy. However, the combined treatment of outliers and policy interventions in forecasting Indian Standard Natural Rubber Grade 20 (ISNR 20) prices remains insufficiently explored.
Aims: To develop a forecasting framework for monthly Indian Standard Natural Rubber Grade 20 (ISNR 20) prices that accounts for both outliers and the July 2015 Rubber Production Incentive Scheme (RPIS) intervention.
Study Design: Time series analysis combining ARIMA-based outlier detection, series cleaning, Bayesian intervention analysis and forecasting.
Place and Duration of Study: Kottayam market, Kerala, India; January 1995 to December 2025.
Sample: 372 monthly observations of ISNR 20 prices (Rs/100 kg); the last 36 months (2023–2025) were held out for testing.
Methodology: The workflow had four steps: ARIMA-based outlier detection, series cleaning, modelling of the RPIS intervention and forecasting.
Results: The price series was non-stationary (ADF = −2.19, P = .49) and became stationary after first differencing (ADF = −6.53, P < .01). Three outliers were detected: a level shift in October 2008 and additive outliers in November 2011 and August 2012. Average post-intervention prices (Rs 13,443) were lower than the counterfactual prices expected without the intervention (Rs 15,816; 95 per cent credible interval 14,538–17,081), an average effect of −15 per cent (95 per cent CrI: −21 to −7.5 per cent), with a posterior probability of an effect of 99.89 per cent. The final ARIMA (1,1,0) model, evaluated on the 36-month hold-out, gave an RMSE of 2626.35 and a MAPE of 11.27 per cent.
Conclusion: Outlier handling and intervention modelling can be combined in one framework for agricultural price forecasting, and treating outliers before intervention modelling gives a more stable baseline. Prices after July 2015 were lower than those implied by the pre-July 2015 dynamics, but this should be interpreted as an association rather than a causal effect of RPIS alone.
Keywords: Outlier detection, natural rubber, price forecasting, intervention analysis, Bayesian analysis, level shift, time series, rubber production incentive scheme