nep-for New Economics Papers
on Forecasting
Issue of 2022‒01‒17
six papers chosen by
Rob J Hyndman
Monash University

  1. Real-Time Forecasting with a (Standard) Mixed-Frequency VAR During a Pandemic By Frank Schorfheide; Dongho Song
  2. "An application of deep learning for exchange rate forecasting". By Oscar Claveria; Enric Monte; Petar Soric; Salvador Torra
  3. Multivariate Realized Volatility Forecasting with Graph Neural Network By Qinkai Chen; Christian-Yann Robert
  4. Volatility and Dependence Models with Applications to U.S. Equity Markets By Pan, Jingwei
  5. NumHTML: Numeric-Oriented Hierarchical Transformer Model for Multi-task Financial Forecasting By Linyi Yang; Jiazheng Li; Ruihai Dong; Yue Zhang; Barry Smyth
  6. Machine Learning for Predicting Stock Return Volatility By Damir Filipović; Amir Khalilzadeh

  1. By: Frank Schorfheide; Dongho Song
    Abstract: We resuscitated the mixed-frequency vector autoregression (MF-VAR) developed in Schorfheide and Song (2015, JBES) to generate macroeconomic forecasts for the U.S. during the COVID-19 pandemic in real time. The model combines eleven time series observed at two frequencies: quarterly and monthly. We deliberately did not modify the model specification in view of the COVID-19 outbreak, except for the exclusion of crisis observations from the estimation sample. We compare the MF-VAR forecasts to the median forecast from the Survey of Professional Forecasters (SPF). While the MF-VAR performed poorly during 2020:Q2, subsequent forecasts were at par with the SPF forecasts. We show that excluding a few months of extreme observations is a promising way of handling VAR estimation going forward, as an alternative of a sophisticated modeling of outliers.
    JEL: C11 C32 C53
    Date: 2021–12
  2. By: Oscar Claveria (AQR IREA, University of Barcelona (UB). Department of Econometrics, Statistics and Applied Economics, University of Barcelona, Diagonal 690, 08034 Barcelona, Spain. Tel.: +34-934021825); Enric Monte (Department of Signal Theory and Communications, Polytechnic University of Catalunya (UPC)); Petar Soric (Faculty of Economics & Business University of Zagreb.); Salvador Torra (Riskcenter–IREA, University of Barcelona (UB).)
    Abstract: This paper examines the performance of several state-of-the-art deep learning techniques for exchange rate forecasting (deep feedforward network, convolutional network and a long short-term memory). On the one hand, the configuration of the different architectures is clearly detailed, as well as the tuning of the parameters and the regularisation techniques used to avoid overfitting. On the other hand, we design an out-of-sample forecasting experiment and evaluate the accuracy of three different deep neural networks to predict the US/UK foreign exchange rate in the days after the Brexit took effect. Of the three configurations, we obtain the best results with the deep feedforward architecture. When comparing the deep learning networks to time-series models used as a benchmark, the obtained results are highly dependent on the specific topology used in each case. Thus, although the three architectures generate more accurate predictions than the time-series models, the results vary considerably depending on the specific topology. These results hint at the potential of deep learning techniques, but they also highlight the importance of properly configuring, implementing and selecting the different topologies.
    Keywords: Forecasting, Exchange rates, Deep learning, Deep neural networks, Convolutional networks, Long short-term memory. JEL classification: C45, C58, E47, F31, G17.
    Date: 2022–01
  3. By: Qinkai Chen; Christian-Yann Robert
    Abstract: The existing publications demonstrate that the limit order book data is useful in predicting short-term volatility in stock markets. Since stocks are not independent, changes on one stock can also impact other related stocks. In this paper, we are interested in forecasting short-term realized volatility in a multivariate approach based on limit order book data and relational data. To achieve this goal, we introduce Graph Transformer Network for Volatility Forecasting. The model allows to combine limit order book features and an unlimited number of temporal and cross-sectional relations from different sources. Through experiments based on about 500 stocks from S&P 500 index, we find a better performance for our model than for other benchmarks.
    Date: 2021–12
  4. By: Pan, Jingwei
    Abstract: The dissertation consists of three studies concerning the research fields of evaluating volatility and correlation forecasts as well as modeling of tail dependence. Based on theoretical discussions and empirical studies the methods for modeling the time-varying volatilities and dependence for the financial market data are evaluated. The first study evaluates the volatility forecasts with the basic generalized conditional autoregressive heteroskedasticity (GARCH) model and its asymmetric extensions. The concepts of loss function and model confidence set (MCS) are introduced. The realized volatility is used as benchmark. The main results of Brownlees et al. (2011) can be confirmed and extended. In particular, the one-step forecasts achieve significantly lower average losses than the multi-step forecasts in times of crises. The difference between the one-step and the multi-step forecasts in pre-crisis times is relatively small. The evaluation results demonstrate the strong forecasting performance of the asymmetric model variants. The second study evaluates the multivariate correlation forecasts. The Baba-Engle-Kraft-Kroner (BEKK) model of Engle and Kroner (1995) is compared with the dynamic conditional correlation (DCC) model of Engle (2002). Using a two-stage estimation method, the DCC model is well suited for large correlation matrices. In contrast, the more flexible BEKK model suffers from the curse of dimensionality. The evaluation is based on the class of asymmetric loss functions proposed by Komunjer and Owyang (2012). The results show that the BEKK model cannot better predict the correlations than the simpler DCC model in the trivariate system. Therefore, the application of the DCC model appears to be superior. The third study leads to a flexible approach which separates the univariate marginal distributions from the joint distribution. The different copula functions are presented and the corresponding tail dependence is calculated. The empirical analysis compares different copula functions with a non-parametric approach and three time-dependent approaches. The results show noticeable reactions of tail dependence to the major financial market events. In addition, the lower tail dependence dominates over time. This can be interpreted in a way that joint losses occur more frequently than joint gains.
    Date: 2021
  5. By: Linyi Yang; Jiazheng Li; Ruihai Dong; Yue Zhang; Barry Smyth
    Abstract: Financial forecasting has been an important and active area of machine learning research because of the challenges it presents and the potential rewards that even minor improvements in prediction accuracy or forecasting may entail. Traditionally, financial forecasting has heavily relied on quantitative indicators and metrics derived from structured financial statements. Earnings conference call data, including text and audio, is an important source of unstructured data that has been used for various prediction tasks using deep earning and related approaches. However, current deep learning-based methods are limited in the way that they deal with numeric data; numbers are typically treated as plain-text tokens without taking advantage of their underlying numeric structure. This paper describes a numeric-oriented hierarchical transformer model to predict stock returns, and financial risk using multi-modal aligned earnings calls data by taking advantage of the different categories of numbers (monetary, temporal, percentages etc.) and their magnitude. We present the results of a comprehensive evaluation of NumHTML against several state-of-the-art baselines using a real-world publicly available dataset. The results indicate that NumHTML significantly outperforms the current state-of-the-art across a variety of evaluation metrics and that it has the potential to offer significant financial gains in a practical trading context.
    Date: 2022–01
  6. By: Damir Filipović (Ecole Polytechnique Fédérale de Lausanne; Swiss Finance Institute); Amir Khalilzadeh (Ecole Polytechnique Fédérale de Lausanne)
    Abstract: We use machine learning methods to predict stock return volatility. Our out-of-sample prediction of realised volatility for a large cross-section of US stocks over the sample period from 1992 to 2016 is on average 44.1% against the actual realised volatility of 43.8% with an R2 being as high as double the ones reported in the literature. We further show that machine learning methods can capture the stylized facts about volatility without relying on any assumption about the distribution of stock returns. Finally, we show that our long short-term memory model outperforms other models by properly carrying information from the past predictor values.
    Keywords: Volatility Prediction, Volatility Clustering, LSTM, Neural Networks, Regression Trees.
    JEL: C51 C52 C53 C58 G17
    Date: 2021–12

This nep-for issue is ©2022 by Rob J Hyndman. It is provided as is without any express or implied warranty. It may be freely redistributed in whole or in part for any purpose. If distributed in part, please include this notice.
General information on the NEP project can be found at For comments please write to the director of NEP, Marco Novarese at <>. Put “NEP” in the subject, otherwise your mail may be rejected.
NEP’s infrastructure is sponsored by the School of Economics and Finance of Massey University in New Zealand.