A fast and scalable ensemble of global models with long memory and data partitioning for the M5 forecasting competition

This work presents key insights on the model development strategies used in our cross-learning-based retail demand forecast framework. The proposed framework outperforms state-of-the-art univariate models in the time series forecasting literature. It has achieved 17th position in the accuracy track of the M5 forecasting competition, which is among the top 1% of solutions.     

Are voters rational?

We test whether a voter's decision to cast a vote depends on its probability of affecting the election outcome. Using exogenous variation arising at population cutoffs determining council sizes in Finnish municipal elections, we show that larger council size increases both pivotal probabilities and turnout. These effects are statistically significant, fairly large and robust. Finally, we use a novel instrumental variables design to show that the jumps in the pivotal probabilities are the likely candidate for explaining the increase in turnout, rather than the other observed simultaneous jumps at the council size cutoffs. Moreover, our results indicate that turnout responds only to within-party pivotal probabilities, perhaps because they are more salient to the voters than the between-party ones.     

A hybrid machine learning model for forecasting a billing period’s peak electric load days

Many models have been studied for forecasting the peak electric load, but studies focusing on forecasting peak electric load days for a billing period are scarce. This focus is highly relevant to consumers, as their electricity costs are determined based not only on total consumption, but also on the peak load required during a period. Forecasting these peak days accurately allows demand response actions to be planned and executed efficiently in order to mitigate these peaks and their associated costs. We propose a hybrid model based on ARIMA, logistic regression and artificial neural networks models. This hybrid model evaluates the individual results of these statistical and machine learning models in order to forecast whether a given day will be a peak load day for the billing period. The proposed model predicted 70% (40/57) of actual peak load days accurately and revealed potential savings of approximately USD $80,000 for an American university during a one-year testing period.     

Combining forecasts for universally optimal performance

There are two potential directions of forecast combination: combining for adaptation and combining for improvement. The former direction targets the performance of the best forecaster, while the latter attempts to combine forecasts to improve on the best forecaster. It is often useful to infer which goal is more appropriate so that a suitable combination method may be used. This paper proposes an AI-AFTER approach that can not only determine the appropriate goal of forecast combination but also intelligently combine the forecasts to automatically achieve the proper goal. As a result of this approach, the combined forecasts from AI-AFTER perform well universally in both adaptation and improvement scenarios. The proposed forecasting approach is implemented in our R package AIafter, which is available at https://github.com/weiqian1/AIafter.     

A multivariate distance nonlinear causality test based on partial distance correlation: a machine learning application to energy futures

This paper proposes a multivariate distance nonlinear causality test (MDNC) using the partial distance correlation in a time series framework. Partial distance correlation as an extension of the Brownian distance correlation calculates the distance correlation between random vectors X and Y controlling for a random vector Z. Our test can detect nonlinear lagged relationships between time series, and when integrated with machine learning methods it can improve the forecasting power. We apply our method as a feature selection procedure and combine it with the support vector machine and random forests algorithms to study the forecast of the main energy financial time series (oil, coal, and natural gas futures). It shows substantial improvement in forecasting the fuel energy time series in comparison to the classical Granger causality method in time series.     

Limitations of portfolio diversification through fat tails of the return Distributions: Some empirical evidence

This study investigates the level of risk due to fat tails of the return distribution and the changes of tail fatness (TF) through portfolio diversification. TF is not eliminated through portfolio diversification, and, interestingly, the positive tail has declining fatness until a certain level is reached, while the negative tail has rising fatness. This indicates that fat tails are highly relevant to common factors on systematic risk and that the relevance of common factors is higher for the negative tail compared to the positive tail. In the portfolio diversification effect, the declining fatness of the positive tail further reduces risk, but the rising fatness of the negative tail does not contribute to this effect. The asymmetry between the fatness of the positive and negative tails in the return distribution corresponds to the asymmetry of the trade-off relationship between loss avoidance and profit sacrifice that is expected as a consequence of portfolio diversification. Investors use portfolio diversification to reduce their risk of suffering high losses, but following this strategy means sacrificing high-profit potential. Our study provides empirical confirmation for the practical limitation of portfolio diversification and explains why investors with diversified portfolios suffer high losses from market crashes. An examination of the Northeast Asian stock markets of China, Japan, Korea, and Taiwan show identical results.     

Modelling demand for lotto using a novel method of correcting for endogeneity

Modelling lottery sales as a function of the mean, standard deviation and skewness of the probability distribution of returns potentially gives insights into how the design of a game could be modified to maximise net revenue. But use of OLS is problematic because the level of sales itself affects values of the moments (and insufficient instruments are available for IV regression). We draw on the concept of a rational expectations equilibrium, developing a new regression model which corrects for endogeneity where the causal impact of the dependent variable on the right-hand side variables is deterministic. We apply the model to data on lotto sales from Spain. Using the Spanish data, we show that results provide more reliable guidance to lottery agencies because accounting for endogeneity leads to significantly different results from OLS and these results have superior performance in out-of-sample forecasting of sales. More generally, results prove consistent with the Friedman-Savage explanation of why people buy lottery tickets and with evidence from racetrack data that ‘bettors love skewness’.     

Predicting bank insolvencies using machine learning techniques

Proactively monitoring and assessing the economic health of financial institutions has always been the cornerstone of supervisory authorities. In this work, we employ a series of modeling techniques to predict bank insolvencies on a sample of US-based financial institutions. Our empirical results indicate that the method of Random Forests (RF) has a superior out-of-sample and out-of-time predictive performance, with Neural Networks also performing almost equally well as RF in out-of-time samples. These conclusions are drawn not only by comparison with broadly used bank failure models, such as Logistic, but also by comparison with other advanced machine learning techniques. Furthermore, our results illustrate that in the CAMELS evaluation framework, metrics related to earnings and capital constitute the factors with higher marginal contribution to the prediction of bank failures. Finally, we assess the generalization of our model by providing a case study to a sample of major European banks.     

On the existence of the ex post symmetric random entry model

This paper studies symmetry among countably infinitely many agents who randomly enter into a stochastic process, one for each period. Upon entry, they observe only the current period signal and try to draw inference about the underlying state governing the stochastic process. We show that there exist random entry models under which agents are ex post symmetric. That is, all agents have identical posterior belief about the underlying states, although they are not ex ante symmetric. The form of the posterior belief is uniquely pinned down by ex post symmetry and a stationarity condition. Our results provide a common prior foundation for the model studied in Liu and Skrzypacz (2014).     

Ex-post incentive compatible and individually rational assignments in housing markets with interdependent values

This paper presents a model for housing markets with interdependent values. We introduce private information on the quality of a house (i.e., high or low), which is known only to the initial owner. Interdependency means that the ex-post preference of an agent depends on the private information of the other agents with regard to the quality of houses. We prove that on a domain satisfying a richness condition, the no-trade rule is the only rule that satisfies ex-post incentive compatibility and ex-post individual rationality.