Just imagine you ordered a cheeseburger. Then low and behold, a salad is brought to your table. The probability of such a thing happening is not zero, especially in a busy restaurant, mix ups happen. It’s happened to all of us. Spare a thought for central banks over the past few years, who repeatedly got served salad, when looking at how their inflation forecasts panned out. There have been articles in the press which have tried to dissect precisely what wrong for central bank forecasts recently (see Central banks rethink forecasting after failures on inflation – FT 27 Dec 2023 and Ben Bernanke to lead Bank of England review into forecasting – FT 28 Jul 2023). In this article, I’ll try to summarize some of the points these articles make, whilst later adding some of my own suggestions on how inflation forecasting can be improved.
Shocking stuff when it comes to inflation
Notably, the main part of the issue with central banks inflation forecasts was not seeing how a plethora of shocks, whether it was COVID to the Ukraine invasion could push up inflation as significantly as it did. Then there is the whole team transitory debate about whether inflation’s rise was purely a result of these shocks, or would be more long lasting (see Why team transitory is still wrong – FT 27 Dec 2023). Then there’s been folks like Larry Summers, who suggested that an extended period of high unemployment was necessary for inflation to fade (see Larry Summers Says US Needs 5% Jobless Rate for Five Years to Ease Inflation – Bloomberg 22 Jun 2022). Admittedly, the former Treasury Secretary is not a central banker, but his view has not been unique. On the other side of the argument, have been economists such as Claudia Sahm who have argued against using the Phillips curve ie. the trade off between inflation and unemployment (see Claudia Sahm: it’s clear now who was right – FT 07 Dec 2023). Sahm suggested that inflation could come down without a large increase in unemployment, a view which does seem to have materialised.
We must first acknowledge that the shocks were not predictable. Would anyone have been able to predict a global pandemic in December 2019 that would envelope the world weeks later? It seems doubtful. As for Russia’s invasion of Ukraine, probably another one where hindsight was invaluable, in particular when it came to predicting its impact on inflation? One important point to note is that inflation started to rise only in 2021 in both the United States and Eurozone, many months after the initial shock of COVID. Arguably the shock of the Russia invasion was far quicker in pushing up inflation, in particular on the Eurozone, which was directly impacted by the gas supply from Russia. However, once these shocks had happened, perhaps forecasters were not “off the hook” when it came to revising their forecasts.
If we think about inflation it is driven by supply and demand of goods and services. From a policymaker perspective, if you believe that inflation in a particular period of time is demand driven, changing monetary policy can impact that albeit with a lag. On the other hand, if it’s more supply driven, monetary policy isn’t going to be as effective (unless the Fed can suddenly expand the supply of ships to unblock supply chains and drill for oil…). In practice, it can often be a mixture of the two. Indeed, in the most recent episode of inflation, Why team transitory is still wrong – FT 27 Dec 2023 notes that the reason why inflation has come down recently is a result of both supply chain issues getting resolved and also Fed policy, citing a study from Allianz. It is also challenging to know what would have happened to inflation if central banks did not tighten.
Use more relevant data and better models to improve inflation forecasts
If you are forecasting inflation, whether or not you are policymaker, what can you do to improve your accuracy? One is to try to decipher which variables are most important and focus your work on a simple inflation forecasting model based on those factors. The problem is that the drivers for inflation are not constant and they change. Rewind to 2021-22, we found that are model forecasts were being impacted by variables related to supply chain disruptions related to COVID and those showing the easing of monetary policy. As supply chains have eased, variables related to them have been weighted less by our inflation forecasting models, which seems intuitive.
Rather than trying to use simple models that overemphasis “easy to understand” relationships between inflation and the rest of the economy, which may or may not be applicable at a particular point in time, we should expand the dataset which is used by inflation forecasting models. Our dataset should encompass a larger array of potential variables that drive inflation. I’m not talking about ingesting many terabytes of data to train a model, a la ChatGPT, from every part of the internet. Instead, I’m suggesting that we should cast a wider net on the relevant data we use. We can then allow the model to come up with its own judgements of what variables are important at any point in time, rather than forcing it down an avenue based on our own particular set of beliefs, which can introduce bias. If folks want to put their own spin on top of the output of the model that’s fine, but the model output itself should data driven.
We live in a Big Data world. We now have data on the economy that would be unimaginable even a decade ago. Yet there is still an obsession with using something like the Phillips curve which has two variables: inflation and unemployment. Of course traditional macroeconomic datasets are still crucial when it comes to forecasting inflation, whether it’s inflation (and its lags), wages, GDP, consumer prices etc. The key point is that we can augment our traditional variables with alternative data. For example, we have data that monitors supply chains (such as the the NYFed’s Global Supply Chain Pressure Index). We have worldwide data on pollution which can a high frequency proxy for industrial activity. We have text data from the web discussing inflation. We have a massive plethora of survey data related to inflation and relevant market data too. It isn’t easy to curate such a relevant dataset for inflation forecasting, and it can vary significantly between countries. However, the raw data is there if we look hard enough, and can be utilised if we spend enough time to clean and prepare it.
On the modelling side, whilst much of the focus has been on the use of models for natural language processing (eg. ChatGPT and LLMs more broadly) we do have machine learning models that can deal with both nonlinear and linear relationships when forecasting time series. They can deal with large datasets, even if many of the variables might be correlated, where an OLS would not be appropriate. Our experience in forecasting inflation at Turnleaf Analytics has been that whilst the modelling is important, probably even more crucial is the dataset you ingest. A fancy model will not help you if you do not have a relevant and good quality dataset. Of course, the flipside of using a large dataset in tandem with a more complicated model is that whilst the accuracy is higher, it can be trickier to explain the output (although, there are techniques that can help to open up the black box).
You could argue that whilst central banks were indeed surprised by the extent of inflation, inflation has indeed come down to more normal levels, even if it isn’t yet within their target. Of course forecasting inflation isn’t easy, as central banks and many others in the market have found. Whichever technique we use, there will always be surprises. However, by utilising a broader set of variables and machine learning techniques, we can get more accurate inflation forecasts, and reduce the size of the surprises. What is also clear is that even if inflation is normalising, we still need to keep an eye on it, in particular in emerging markets where inflation has always been an issue, and for that accurate forecasts are crucial. The environment of ZIRP, is behind us, and we are back to a normal world, with a modicum of inflation volatility.
Final question, who wants a cheeseburger, instead of salad. My answer: I want a cheeseburger.