Machine learning in economics

Abstract

Given an arbitrary black-box strategy for the Iterated Prisoner’s Dilemma game, it is often difficult to gauge to which extent it can be exploited by other strategies. In the presence of imperfect public monitoring and resulting observation errors, deriving a theoretical solution is even more time-consuming. However, for any strategy the reinforcement learning algorithm Q-Learning can construct a best response in the limit case. In this article I present and discuss several improvements to the Q-Learning algorithm, allowing for an easy numerical measure of the exploitability of a given strategy. Additionally, I give a detailed introduction to reinforcement learning aimed at economists.

With increasing demand for wireless data and new requirements to uphold “net neutrality,” internet service providers try new methods to ensure their profits. Sponsored content, the archetype of “Zero-Rating,” allows the content provider to pay for the accrued traffic instead of the consumer.

This paper shows, using a theoretical model, that allowing sponsored content has ambiguous results both on infrastructure investments and net welfare.

Sponsored content may be used by the ISP to internalize additional profits of the content provider and thus motivate higher infrastructure investments, especially given very high network costs. However, given a sufficiently high profitability of the content provider, the internet service provider might be more interested in optimizing the income stream from the content provider than in the satisfaction of the consumer. This not only decreases effective network capacity but might also lead to a decrease in net welfare.

Instead of relying solely on data of a single time series it is possible to use information of parallel, similar time series to improve prediction quality. Our data set consists of microeconomic data of daily store deposits from a large number of different stores. We analyze how prediction performance regarding a given store can be increased by using data from other stores. First we compare several machine learning methods, including Elastic Nets, Partial Least Squares, Generalized Additive Models, Random Forests, Gradient Boosting and Neural Networks using only data of a single time series. Afterwards we show that Random Forests are able to better utilize parallel time series data compared to Partial Least Squares. Using near-time data of parallel time series is highly beneficial for prediction performance. To allow a fair comparison between different machine learning methods, we present a novel hyper-parameter optimization technique using a regression tree. It enables a fast and flexible determination of optimal parameters for a given method.