|
on Forecasting |
| By: | MacLachlan, Matthew; Chelius, Carolyn; Short, Gianna |
| Abstract: | This technical bulletin describes a time-series-based approach for forecasting food prices that includes prediction intervals to communicate uncertainty. The performance of forecasts created with this approach was compared to that of previously published USDA, Economic Research Service (ERS) Food Price Outlook (FPO) forecast ranges. The methods in this new approach are intended to be used in FPO data releases that provide monthly forecasts of annual food price changes and may also prove useful in other forecasting endeavors. The new approach used an autoregressive integrated moving average (ARIMA) model that was selected based on performance (information loss), generating a more accurate forecast than previously used methods as measured by root-mean-square errors. With the parameter estimates and estimated error distribution from the optimal ARIMA model, Monte Carlo simulations are used to develop prediction intervals, which reflect uncertainty about future food prices. These prediction intervals more often included the actual annual price changes than the archived fore-cast ranges. On average, the prediction intervals also included the actual annual price change earlier in the forecasting process. These properties generally held whether we used a higher (95 percent) or lower (90 percent) confidence level. The use of standardized econometric models and model selection also allowed for the inclusion of data not currently included in FPO. The methods easily tested whether including external variables improved forecast accuracy or could be used to create new forecasts. This report considered new price change forecasts of apples, seafood, and limited-service restaurants in 2020 and the potential forecast performance improvement from incorporating futures prices as case studies. |
| Keywords: | Consumer/Household Economics, Demand and Price Analysis, Research Methods/ Statistical Methods, Risk and Uncertainty |
| Date: | 2022–08–18 |
| URL: | http://d.repec.org/n?u=RePEc:ags:usdami:327370&r=for |
| By: | Stephen G. Hall (Leicester University, Bank of Greece, and Pretoria University); George S. Tavlas (Bank of Greece and the Hoover Institution, Stanford University); Yongli Wang (University of Birmingham) |
| Abstract: | This paper considers the problem of forecasting inflation in the United States, the euro area and the United Kingdom in the presence of possible structural breaks and changing parameters. We examine a range of moving window techniques that have been proposed in the literature. We extend previous work by considering factor models using principal components and dynamic factors. We then consider the use of forecast combinations with time-varying weights. Our basic finding is that moving windows do not produce a clear benefit to forecasting. Time-varying combination of forecasts does produce a substantial improvement in forecasting accuracy. |
| Keywords: | forecast combinations, structural breaks, rolling windows, dynamic factor models, Kalman filter |
| JEL: | C52 C53 |
| Date: | 2022–10 |
| URL: | http://d.repec.org/n?u=RePEc:bir:birmec:22-12&r=for |
| By: | Fernando Moreno-Pino; Stefan Zohren |
| Abstract: | Volatility forecasts play a central role among equity risk measures. Besides traditional statistical models, modern forecasting techniques, based on machine learning, can readily be employed when treating volatility as a univariate, daily time-series. However, econometric studies have shown that increasing the number of daily observations with high-frequency intraday data helps to improve predictions. In this work, we propose DeepVol, a model based on Dilated Causal Convolutions to forecast day-ahead volatility by using high-frequency data. We show that the dilated convolutional filters are ideally suited to extract relevant information from intraday financial data, thereby naturally mimicking (via a data-driven approach) the econometric models which incorporate realised measures of volatility into the forecast. This allows us to take advantage of the abundance of intraday observations, helping us to avoid the limitations of models that use daily data, such as model misspecification or manually designed handcrafted features, whose devise involves optimising the trade-off between accuracy and computational efficiency and makes models prone to lack of adaptation into changing circumstances. In our analysis, we use two years of intraday data from NASDAQ-100 to evaluate DeepVol's performance. The reported empirical results suggest that the proposed deep learning-based approach learns global features from high-frequency data, achieving more accurate predictions than traditional methodologies, yielding to more appropriate risk measures. |
| Date: | 2022–09 |
| URL: | http://d.repec.org/n?u=RePEc:arx:papers:2210.04797&r=for |