The spot prices of base metals, which are used for constructing infrastructure and various types of products, can be incredibly difficult to predict. The prices are stochastic and highly non-linear as they are influenced by many factors with complex relationships. They are also affected by macro-economic situations such as currency exchange rates and government policy changes, such as increase in import tax rates.
Four Elements Capital, a specialised macro and commodity quantitative asset manager, wanted to find out if artificial intelligence (AI) and machine learning (ML) techniques could be used to improve the accuracy of spot price predictions for base metals. Such techniques will leverage not only market data (trading prices and volumes), macroeconomic data, supply and demand data and third-party estimations but also alternative data sources such as news, specialised forums, reports, social media and demand forecasts.
The firm is promoting collaborative research through its Alphien product, which is an open platform to conduct financial research providing the data and the computing power to allow academics to work on financial projects. Its managing director, Lionel Semonin, believes that the ability to integrate AI successfully into the financial domain will position Singapore as the global financial centre of the future.
Taking a step towards this goal, Four Elements worked with AI Singapore through the 100 Experiments (100E) programme to develop an ML framework for price forecasting on the Alphien platform. The framework consisted of three components. The first was the classification of the direction of price movement using an ensemble of ML and deep learning models. The second was a regression of price predictions using a deep learning model. The third was a filter mechanism to align the outputs from the above.
The team also developed new indicators to extract intelligence from alternative data sources relevant to base metal trading, such as news and analyst reports, to derive short-term price movement indicators.
When the source code was deployed on the Alphien server for testing, it was found that the ML framework outperformed traditional linear model benchmarks in its accuracy for out-of-sample forecasts for one, three, five and 10-day horizons.
“There is potential for the ML framework to be spun off and scaled to help financial institutions, business owners and government entities better manage their risk exposure,” said Semonin. “Implementing the AI solution in the financial world will help strengthen and develop Singapore as a leading fintech hub.”