Inflation changes the cost of getting to work and the cost of what is on the table. Anticipating those changes is hard, and one line of research asks whether machine learning can do it better than reacting after the fact. This post looks at a 2024 study by Harsh Gupta, R. Aafrein, Rashmi Kumari, Swapnil Rajput, and Nupur Puri, published in the International Journal of Scientific Research in Science and Technology. The study is not specific to the Philippines, but the question it asks applies here.
The setup
The authors collected historical records for 14 commodities from financial databases, chosen for their market significance and trading volume. The series ran from January 2012 to December 2021 and were standardized before modeling. Alongside the prices, the study considers the economic indicators that move them: supply and demand, geopolitical events, and broader macroeconomic trends.
Eight methods were compared. Four are general-purpose machine learning models: support vector machines, decision tree regressors, k-nearest neighbors regression, and linear regression. Three are classical multivariate time series models: vector autoregression (VAR), VARMA, and VARIMA. The last is a long short-term memory (LSTM) neural network. Each was scored on standard regression metrics for precision, error, and goodness of fit.
What they found
The decision tree regressor explained nearly 99 percent of the variance in commodity prices. Linear regression scored close to perfect, which the authors read as a warning sign of overfitting rather than a win. Support vector machines and the LSTM had moderate errors with good explanatory power. The k-nearest neighbors regressor and the VAR family had the highest errors and the weakest fits.
What they recommend
The authors suggest that people making trading, investment, or risk decisions lean on the stronger models, the decision tree regressor and the LSTM, and integrate them into systems that update as new prices arrive. They also stress two cautions. Models that look too good, like the linear regression here, need strict validation to rule out overfitting. And the quality of the forecasts depends on preprocessing and feature engineering as much as on the choice of algorithm.