Nowcasting with Dynamic Factor Model and Real-Time Vintage Data: A financial market actor's perspective

Abstract

We develop and examine a dynamic factor nowcasting model (DFM) from the perspective of a financial market participant. The first point of analysis is the examination of its performance. Unlike other papers, we evaluate with daily frequency so that the performance metric reflects a continuous nowcasting signal. Secondly, we examine the effect of using real-time vintage data which avoids look-ahead bias, compared to the common practice pseudo real-time vintage data. We conclude that the DFM outperforms a simple benchmark AR model and an alternative factor model approach. However, it is unsuccessful in incorporating newly released information to improve its estimate in the second half of a quarter. We also conclude that using pseudo real-time data sets may be misleading during times of high volatility due to economic variables being revised, or when using a DFM of less than 4 factors. In general, however, the differences in performance between the data approaches are small when modeling with a DFM, indicating that using pseudo data sets is a reasonable approach to tackle the issue of short supply of vintage data.

Popular Abstract

GDP forecasting from a financial actor’s perspective
Accurate economic short-term forecasting (a.k.a. Nowcasting) provides investors with competitive advantages by informing investment decisions. This thesis evaluates a common forecasting approach from the perspective of an investor by making daily GDP forecasts and using two types of data, one that is more complex than the other. We conclude that a Dynamic Factor Model outperforms benchmark models and that using simplified data sets is a good approximation when testing models, under some conditions.
Successfully predicting the short-term future is a valuable skill. This type of forecast, called Nowcasting, originated in meteorology and has since spread to other fields. Within economics and finance, central banks and investors may use accurate nowcasts to design policy and strategy. It does not take much imagination to see that an investor with superior knowledge of the short-term economic development has a competitive advantage. This is the perspective taken in this thesis, where a nowcasting model is evaluated from the viewpoint of an actor on the financial markets.
The model, called Dynamic Factor Model (DFM), uses 128 economic time series such as Industrial Production Index, Civilian Unemployment Rate, and Housing Starts. These series are fed into the model, after some processing, with the aim of predicting the annualized quarterly GDP growth rate. Between 2005 and 2019, our test interval, we estimate GDP daily. A challenge that one faces when trying to solve this problem is missing data. Imagine that we want to get a GDP estimate for the 5th of January 2010. Industrial Production Index will then only be available up until November 2009; thus, we are missing the December and January figures which are needed to estimate GDP. Using a so-called Kalman Filter, we estimate the December and January figures for the IP index as well as missing figures in other series. These figures are then used to estimate GDP.
What distinguishes this thesis from earlier work is mainly the way the model is evaluated. As financial actors may make investment decisions at any given time, not just once a month or quarter, they need to have frequently updated information at their disposal. This led us to evaluate the model daily. We also use two types of data sets: a more advanced, which is seldom used in the nowcasting field, and a simpler data set that tries to mimic the more complex one. This last one is frequently used in literature.
The resulting model proved to be decent for financial market actors wanting to use Nowcasting in their investment strategy. First, it performed better than our benchmark models tested on the same data set and interval. Secondly, it is easy to adapt it to nowcast other variables than GDP, such as inflation, unemployment rate, etc. However, the model’s inability to improve its estimate as more economic data is made available is a worrying sign. As we get further into a quarter, more relevant data is made available to the model, e.g. on the 15th of January the November figures for the IP index are made available, giving the model one more data point to work with. We, therefore, expect the model to improve as days pass. However, improvement is often limited to the first month of the quarter. We also saw that using the simplified data set is a good approximation, under certain conditions. During times of high volatility, however, we saw that it could produce misleading results. In summation, under some conditions, developing a Dynamic Factor Model and testing it using a simplified data set appears to be a good alternative for the financial market actor.

• author: Östlund, Filip and Attar, Marcel
• supervisor: Erik Lindström
• organization: Mathematical Statistics
• course: FMSM01 20201
• year: 2020
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: Nowcasting, Macroeconomic Prediction, Dynamic Factor Model, DFM, Pseudo Real-Time Vintage data, U.S. GDP Growth Rate, Financial Market Actor
• publication/series: Master's Theses in Mathematical Sciences
• report number: LUTFMS-3395-2020
• ISSN: 1404-6342
• other publication id: 2020:E57
• language: English
• id: 9023901
• date added to LUP: 2020-07-03 16:28:26
• date last changed: 2021-06-04 18:34:05