Speevr logo
Report

Inflation | Can machine learning improve forecasting?

Table of Contents

Announcement: Amazon's Chief Economist, Pat Bajari, is seeking to hire more PhD economists for the firm's new venture in AI consulting services. We are happy to connect between potential candidates with Pat and his colleagues.

Amazon is hiring more PhD economists of new AI consultancy service

Amazon likely employs more PhD economists for research than the San Francisco and New York Feds combined. Clients inevitably end up acquiring cloud services where, in our opinion, Amazon still offers the best customer support. Even for remote cloud servers, price is not always the ultimate determining factor.

Three alternative inflation views and time horizons

Our current inflation outlook is primarily influenced by several sources with varying time horizons. These include Jan Hatzius and his team at Goldman SachsJason Furman at Harvard, and Alan Brazil of SOM Macro. While there are others who have contributed to enriching our understanding, these three sources provide a comprehensive overview of the US inflation picture and potential risks.

A review of recent academic literature on inflation forecasting models

I conducted a brief search for recent academic paper proposals on alternative models to improve inflation forecasts. The results were somewhat underwhelming, and nothing particularly stood out among the peer-reviewed studies published in the past three years. As we stated from the outset in 2021, inflation is a complex topic and something of an enigma, even among the brightest academic scholars.

Underwhelming, but here's what we found… 

Here are a few observations that may be useful based on our review on the recent academic literature and discussions with several market participants:

Inconsistent understanding of what is inflation 

Common measures of inflation, such as CPI, PCE, or the GDP deflator, do not necessarily align with how most ordinary people or even market professionals perceive inflation, i.e., changes in the aggregate price level. Some individuals view inflation as a mental or behavioral process that drives prices. Not only is this an abstract and ill-defined concept that's difficult to prove or disprove, but in some instances, it merely amounts to what economists refer to as long-term inflation expectations.

Easier to stick with common inflation measures

For scientific purposes, it is advisable to stick with the well-defined measures of inflation mentioned until a new era in macroeconomics emerges that significantly improves on existing (imperfect) concepts and theories.

Mostly machine learning applied to inflation forecasting

A majority of recent studies aimed to leverage big data and neural networks (machine learning) to forecast inflation. While these alternative methods have shown better predictive capabilities for the 2021-22 surge in inflation, it remains unclear whether they would perform equally well in out-of-sample conditions in the future. Based on what we know about machine learning, it is highly unlikely that, due to the statistical nature of economic time-series, such computational methods can identify early paradigm shifts in low-frequency data. Therefore, in this situation, simple mathematical models are preferable to non-transparent black-box outputs, even if they are imperfect.

The interim steps in machine learning models offer more clues than the final outputs

There may be scope to utilize machine learning for gaining insights on inflation by conducting meta-studies using a consistent approach for several economies. Also, the interim steps in the model may hold valuable information which is not captured by final output. For example, if a specific input price consistently reveals a high common weight across various economies before/after an episode of inflation surge/decline.

Traditional data mining technique may serve as a better starting point than the additional complexity of neural networks.

A basic unified conclusion from the latest research

Identifying common features recognized by machines, that are not evident to humans, can serve as a hint to new ways of thinking about a problem. It is very possible that there is a unified (relatively trivial) conclusion from the various machine learning studies of inflation: neural networks can automatically reweight constituent inputs quicker than in official calculations.

For instance, if second-hand cars were assigned a higher constitutive weight earlier, official headline CPI calculations would have reflected it accordingly. As such, the findings may not be especially profound but could suggest predictive powers in better understanding inflation calculation methodology.

It becomes a circular argument if implemented

Unfortunately, most research papers primarily focus on differences between machine learning inflation forecasts and the official figures in a single economy, disregarding potentially unique insights. It is very challenging to escape the boundaries and limitations inherent in working within existing frameworks (inflation measures) through purely statistical means. Otherwise, statisticians would dominate every field of research.

What was the common cause of error amongst that led to the big miss in inflation surge?

To the best of our knowledge, a popular stochastic model used by almost all major central banks, the Dynamic Stochastic General Equilibrium(DSGE) model, was the primary culprit responsible for the poor inflation forecasts of 2021/22. More specifically, it relied on a hard-coded assumption that inflation, as a random process, reverts back to the official policy target of 2%. In other words, the policy tail was wagging the inflation dog.

Cognitive bias from recent history than poor choice of models

Without detailed knowledge of the DSGE model's inner workings, based on our general understanding of stochastic processes, it is highly unlikely that an alternative choice of long-term inflation would have satisfactorily explained or anticipated the surge in inflation of 2021/22. More likely, the DSGE models and a prolonged period of low inflation shaped the thinking of a generation of macroeconomists. Speaking from personal experience, the latter was the prevailing factor for dismissing the early surge in US inflation. For many years, combating deflation was the primary concern of most G7 central banks.

The simplest model for a complex system is the best one

In conclusion, major central banks have not rushed to announce new inflation forecasting models to account for the recent surge in inflation, which is mildly encouraging. Complex systems that do not readily fit a particular mathematical model are best served by simpler models that are adequate for 70%-80% of the data sampled. The remaining shortcomings may be somewhat compensated with a deeper understanding of the assumptions in the basic models and the system being studied (the economy).

Sometimes, consulting industry outsiders without the perceptions and prejudices of subject matter experts is an effective path to a new solution.

Subscribe to receive updates from Speevr Intelligence

Most recent by Speevr Intelligence

report

Share this page

Inflation | Can machine learning improve forecasting?

A brief review of the recent academic literature