Insights
A transparent alternative to neural networks
Many prediction tasks in economics and finance lie beyond the reach of linear regression analysis.
October 2024
Researchers, therefore, often turn to machine learning techniques, such as neural networks, to address these complex dynamics.
Key highlights
Nearly a century ago, Rube Goldberg introduced his famous satirical cartoons of mechanical devices that perform simple tasks in convoluted and impractical ways. For example, to put a stamp on an envelope you sneeze, scaring a dog, tipping a hat rack, breaking glass, spilling water, filling a bucket and so on. Nowadays, our sequences are more sophisticated: For example, we use computers to train parameters and form predictions using deep neural networks. In both cases, we get an outcome from a long chain of steps and the purpose of each step is mostly a mystery.
To be fair, neural networks have achieved amazing feats, even if they are unwieldy to operate and hard to decipher. But in our new research, we show that a direct and interpretable approach to forming a prediction, called relevance-based prediction (RBP), delivered values that were highly similar, and often better at predicting volatility across constantly shifting market regimes (see Exhibit 1).
A neural network forms a prediction by performing many sequences of linear combinations and nonlinear transformations of the original inputs. In doing so, it has the potential to extract nearly all the useful information from a dataset. However, a neural network is difficult to implement and notoriously opaque.
Alternatively, RBP is a model free and theoretically-grounded approach that forms a prediction as a relevance-weighted average of past outcomes. It adapts the selection of observations and predictive variables to each prediction task, which enables RBP to capture complex relationships, like a neural network.
However, unlike a neural network, RBP is remarkably transparent, revealing how each observation and variable contributes to a prediction, and disclosing the reliability of a prediction in advance. This transparency, combined with its ease of implementation, suggests that RBP may serve as a compelling alternative to neural networks.
