Financial forecasting using stochastic models: reference from multi-commodity exchange of India

Non-linear forecasting models, including artificial neural networks, are popularly adopted in financial forecasting. These models require high computational infrastructure and resources for practical training and deployment, unlike linear forecasting models. However, this level of infrastructure and resources is not accessible to most market participants. This paper lays down a systematic approach to build a simplistic and effective forecasting model, which empowers investors to make informed decisions using a minimal computational infrastructure. Natural gas futures traded in the Multi-Commodity Exchange of India were identified as an appropriate asset for the study considering the massive, expected movement in energy consumption pattern in India. We have used data analytics and statistical techniques to identify optimal training strategies and features, which resulted in an accurate linear forecasting model. Data analytics also helped to accurately establish the context of the study with the identification of a positive shift in the sentiments of market participants of the asset, which was duly verified using the substantial change in the valuation of the asset. While formulating the forecasting model, several avenues, including identifying weekly and yearly patterns, introducing seasonality and exogenous variables were explored. The paper concludes that such attempts to reduce external dependency and segregate noise data result into better model performance with minimal computational resource and infrastructure.