Automated Earnings Forecasts: Beat Analysts or Combine and Conquer?
- Ryan Ball and Eric Ghysels
- Management Science, forthcoming
- A version of this paper can be found here
- Want to read our summaries of academic finance papers? Check out our Academic Research Insight category.
What are the research questions?
- Does a richer time-series forecasting model improve performance relative to a basic autoregressive model?
- Can a richer time-series forecasting model beat analysts at short horizons of one fiscal quarter or less? In other words, can we fully automate earnings forecast by making them computer-driven without input from analysts?
- If it cannot, does combining the richer time-series model with analyst forecasts conquer analysts by themselves? In other words, can model-based earnings forecast provide complementary information to analysts and improve their quality of output?
What are the Academic Insights?By adding two firm-level stock return variables and six macroeconomic variables, which are observable at a higher frequency and by using two new econometric developments to deal with larger datasets ( mix data sampling regressions -MIDAS- and a combination forecast model), the authors find the following:
- YES- The MIDAS-combination forecast errors are economically and statistically (36%) smaller than those produced by the autoregression time series model
- YES- The MIDAS-combination forecast errors are economically and statistically (10%) smaller than those produced by the analysts. Additionally, the superiority of the MIDAS-combination model is most pronounced when there is high disagreement among analysts
- YES- The unique combination of analysts and MIDAS-combination forecasts is always better to alternative solutions, which means that the two models are complementary in the information they provide
Why does it matter?The authors apply for the first time the MIDAS forecasting technique to firm-level financial statements measures. They do so by leveling the playing field of information used in this field of study. In fact, they increase the amount and frequency of information used in time series models to a level consistent with information used by analysts. They offer evidence that the process of forecasting firm-level earnings can be automated, but more importantly, it provides evidence that these models can be useful for situations where there is reliable high-frequency data but low analysts coverage ( small firms, emerging markets etc.)
The Most Important Chart from the Paper
Prior studies attribute analysts’ forecast superiority over time-series forecasting models to their access to a large set of firm, industry, and macroeconomic information (an information advantage), which they use to update their forecasts on a daily, weekly or monthly basis (a timing advantage). This study leverages recently developed mixed data sampling (MIDAS) regression methods to synthesize a broad spectrum of high-frequency data to construct forecasts of firm-level earnings. We compare the accuracy of these forecasts to those of analysts at short horizons of one-quarter or less. We find that our MIDAS forecasts are more accurate and have forecast errors that are smaller than analysts’ when forecast dispersion is high and when the firm size is smaller. In addition, we find that combining our MIDAS forecasts with analysts’ forecasts systematically outperforms analysts alone, which indicates that our MIDAS models provide information orthogonal to analysts. Our results provide preliminary support for the potential to automate the process of forecasting firm-level earnings, or other accounting performance measures, on a high-frequency basis.