Forecasting non-linear WPI of manufacture of chemicals and chemical products in India: an MLP approach

Forecasting is an instrument of decision-making that makes predictions or estimates about the future based on historical data. Identifying a suitable strategy for forecasting a time series amongst the classical techniques (e.g., exponential smoothing, Auto-Regressive Integrated Moving Average (ARIMA)), Neural approach, and Support Vector Regression (SVR) - another widely used and popular machine learning-based approach, is challenging. The present work aimed at providing a simple (implementation wise), efficient (forecast accuracy wise), and state-of-art Multi-Layer Perceptron (MLP) approach for some selected macroeconomic indices (Wholesale Price Index - i.e., WPI) in India. We looked at the WPIs with non-linear trends identified using the curve-fit method. It's known that the diverse Indian chemical industry contributes notably to India's economic development. In this work, we analyzed the WPI of seventy-seven commodities/items of the "manufacture of chemicals and chemical products" group in India. We detected the indices having non-linear trends by applying the curve-fit method. The curve-fit approach based on statistical rigor identifies the non-linear WPIs. Twenty-five out of seventy-seven indices exhibits non-linear trends. We developed a forecasting approach employing the MLP for these twenty-five non-linear WPIs. The proposed-MLP optimized by hyperparameter tuning offers high accuracy, prediction reliability, and prediction acceptability for all non-linear WPIs. The forecasting performances of the proposed-MLP compared with regression models (Linear, Quadratic, Cubic, Logarithmic, Exponential), exponential smoothing (Holt linear trend, Holt exponential trend, Holt-Winters), state-of-art Auto-ARIMA, and SVR. The MLP outperformed them all. In terms of Mean Absolute Percentage Error (MAPE), the MLP outperform Linear in 88%, Quadratic in 92%, Cubic in 88%, Logarithmic in 72%, exponential in 88%, Holt Linear in 80%, Holt Exponential in 76%, Holt-Winters in 72%, Auto-ARIMA in 56%, and SVR in 56% of cases. We suggest the application of the proposed approach as an alternative for forecasting these twenty-five non-linear WPIs.