An efficient AI Method for better Economic Forecasting

The fundamental problem with economic forecasting lies in the asynchrony of the available economic data. While many key figures, such as production volumes or sales figures, are usually only published monthly or quarterly, financial markets provide huge amounts of data on a daily basis and in real time. However, conventional forecasting models struggle to combine these mixed data frequencies. They often suffer from the so-called ‘curse of dimensionality’: the more data points are fed in, the slower and more unstable the calculations become. To solve this problem, researchers used a relatively new machine learning approach called ‘reservoir computing’ in a study.

‘Reservoir computing’ as an efficient alternative

This is a special form of training digital neural networks. This is an artificial intelligence approach that mimics the functioning of the human brain. These digital neural networks can, for example, receive and process various economic data and ultimately produce an economic forecast. By comparing this forecast with actual economic developments, the internal parameters of this neural network are continuously optimised during training.
Previous training methods aimed to completely rewire the ‘brain’, i.e. to optimise all parameters of the neural network – a very computationally intensive process. With reservoir computing (RC), however, most of the network's parameters remain untouched during training, while only those at the output level are adjusted. How this works can be illustrated using a pond as an example.

Learning only the wave patterns on the shore

Let's imagine an observer standing at the edge of a pond who wants to use the movements of the water to determine the size of stones thrown into the pond. Conventional training methods for neural networks have attempted to learn the physics of the movement of each individual drop of water, which is a very complex learning process. However, to make good estimates, it is actually sufficient to observe the complex wave patterns on the shore and draw conclusions about the size of the stone based on these. This is exactly what RC attempts to do. Here, most of the digital neurons with purely randomly determined, fixed parameters serve as this ‘pond’ (the reservoir). When we throw data into this network, it echoes back as highly complex patterns, similar to the stone creating complex wave patterns at the water's edge. In RC, only the final output level, which is supposed to translate these patterns into economic forecasts, is trained. Translated into the pond metaphor, the ‘observer’ is trained to correctly interpret only the wave patterns of the water at the shore, rather than studying the entire physics of water movement.

"Unlike other common AI methods, models based on reservoir computing require only a limited amount of data for training. This property is also perfectly suited for macroeconomic applications where only limited data is available," says Prof. Ph.D. Lyudmila Grigoryeva from the School of Economics and Political Science at the University of St. Gallen, who participated in the study. Together with colleagues from other universities, she compared this new approach with the standard macroeconomic models used to date in extensive tests.

A fraction of the computing power for better results

The results of the study are promising: the RC model was able to predict the growth of the US gross domestic product over various periods, matching and often even surpassing established methods. However, the superiority of the RC method was particularly impressive in terms of efficiency. While conventional models required enormous computing capacity for training, the RC approach delivered good results at a fraction of the computing power. The study also shows that these RC models can handle large amounts of input data in economic forecasting without losing accuracy.

KOF experimentally applies new method

As the AI industry is also facing increasing sustainability demands, reservoir computing offers a more resource-efficient alternative for data analysis and a more stable method for evaluating diverse economic data. Lyudmila Grigoryeva and her colleague Giovanni Ballarin have also supported the Nowcasting Lab of the KOF Swiss Economic Institute at ETH Zurich in implementing this new method. Economic forecasts for 17 European countries, including Switzerland, can now be accessed there in real time.

Image source: Adobe Stock / Abdulloh