Commodities | Global | Monetary Policy & Inflation
Inflation Model Explainer
- By Sam van de Schootbrugge
- 2 min read
Summary
- We have developed a framework that uses trends in analysts’ forecasts to predict upcoming CPI data releases more accurately.
- It restates consensus accounting for historical errors in individual forecasters and assigns weights to analysts that reduce the prevalence of common information.
What Are We Forecasting?
The Bureau of Labor Statistics releases US inflation data with a lag of one month. Our model predicts CPI values for each upcoming release. For example, in September 2023, we predict official inflation data for August 2023.
What Data Do We Use?
Each month, up to 60 analysts from global institutions predict upcoming inflation prints. We collect these monthly forecasts dating back to January 2004 from Bloomberg. Alongside historical inflation data, these are the only inputs to our model.
What Are the Benefits to Pooling Forecasts?
Analysts’ forecasts tend to be biased, inefficient, and on average too optimistic. Combining them has been shown to reduce the instability risk associated with reliance on a subset of forecasters and enhance the information content in their forecasts.
As such, consensus forecasts have become an important tool for market participants.
However, a rich set of studies have shown consensus forecasts assign too much weight to analysts’ common information. This makes them less able to adjust to shocks.
We pool all available analyst forecasts for US inflation, minimise their biases and aggregate them to reduce the prevalence of common information. This has made our model up to 20% more accurate than consensus over the last two years.
How Do We Construct Our Model?
We begin by adjusting for biases associated with individual characteristics and incentives. This is because analysts behave predictably. For example, they may be prone to making larger forecast errors in periods of higher or more volatile inflation.
These predictable behaviours allow us to evaluate the accuracy of analysts’ forecasts in any given macroeconomic context. We exploit these trends and adjust their predictions accordingly, before aggregation.
Next, we pool these unbiased forecasts according to a set of weights. Depending on the number of analysts available to be aggregated – which is fewer the earlier we go before the release date – we choose either trimmed or performance weights.
Lastly, to reduce common information in our ensemble model while maintaining the cross-sectional richness offered by the individual forecasters, we correct for biases in the mean forecast by using a set of time-series regressions.
How Do We Perform?
The model is trained over a rolling 15-year window. We estimate its ability to predict inflation, out-of-sample, over the last 55 months since January 2019. Our goal has been to beat consensus, which has proven difficult for most analysts in our sample (Chart 1).
The average analyst has performed 12% worse than consensus at predicting inflation since 2019. Our model has outperformed consensus by 6%, with the largest value-add on headline inflation. There are just two global institutions – out of 53 for headline MoM and 32 for headline YoY – that have posted smaller forecast errors than us over the last 55 months.
Over what has been a challenging timeframe to predict inflation, consensus has been accurate to a first decimal place 29% of the time since 2019. Comparatively, the average analyst has performed worse (26%), while we have performed better (32%).
Lastly, it is not just the average error and hit rate that is important. It is also what can be gleaned from our forecast relative to consensus. We find that if we disagree with consensus:
- Core YoY inflation is 15% more likely to move in our direction.
- Headline YoY inflation is 20% more likely to move in our direction.
- Headline MoM inflation is 40% more likely to move in our direction.
- Core MoM inflation is 55% more likely to move in our direction.
Key Conclusions
Our inflation event forecast model balances the benefits of common knowledge and unique analyst insights to generate predictions that have outperformed almost all global institutions over the last five years.
Moreover, it is more accurate and more informative than consensus, helping us to provide valuable insights into the level and the direction of inflation, whether we predict inflation correctly or not. Its outperformance has been particularly impressive over the last two years.