# Why did economic forecasters get inflations predictions so wrong?

I come from more of a statistics and applied math background, but I am interested in issues of social science. I have been trying to understand why economic forecasters or forecast models seems to have make such bad predictions about inflation over the past few months. Just a caveat, I don't mean to bash on economic forecasters or their models etc., this is just legitimate academic curiosity. I was just looking at some recent articles on this topic, and the explanations seem to be all over the board, without grounding in any common foundation.

Some of these articles might be click-bait, but there is a real question here. I imagine that economists have large forecasting models run by Treasury departments around the world, as well as say Central Banks, and academic departments. These are probably variants of Dynamic Stochastic General Equilibrium (DSGE) models. These models have been refined over years and years with lots of data and additional computing power, etc. In parallel, climateological models have also received lots of investment and support to improve their models since the 1950s. With climate models, the accuracy of the predictions has increased geometrically, if not more, over the past 70 years.

In reading the journalistic pieces I referenced--written by academic economists--there are a lot of references to things like

Slowing the reopening of the economy is commonly cited as a reason that inflation was higher than expected in the second half of 2021. But the rapid reopening of the economy as people were vaccinated in the first half of the year was also cited as a reason for rapid inflation then.

But I am not sure that this is a valid reason? I mean the forecast models don't need to track covid specifics, but if there are obvious imbalances in trade statistics or payrolls or inventory accumulations, then shouldn't the forecasting models still be able to propagate that information into the future? I mean climate forecasting models did not anticipate the reduction in CO2 and other greenhouse gases due to factory closures, but we don't see degraded predictions in climate and weather forecasts.

In the second article reference, Jason Furman, economist suggests how unrealistic predictions made by the Council of Economic Advisors were on unemployment. The problem with this prediction is that they are not just bad, but they are essentially "non-physical" outputs. Here is the quote.

Figure 2 mechanically translates the GDP numbers into unemployment numbers based on the relationship between GDP growth and employment gains described by the Council of Economic Advisers (2009). In the case of the normal multipliers, this results in the economically absurd forecast of a 1.1% unemployment rate in the first quarter of 2021, a sign that something is wrong with this methodology – a topic I will return to.2

It is obvious that an unemployment rate of 1.1% is essentially impossible. In studying numerical methods for partial differential equations, we would call outputs like this "non-physical" in the sense that a numerical PDE solver is giving predictions that are not physically possible. Correcting non-physical solutions has spawned entire literatures on finite volume methods and conservation laws, etc. If a model is giving such crazy outputs, then it is obviously not respecting some basic conservation laws, right?

So I am just trying to understand what are the issues with current forecasting models and methodologies that lead to these types of forecasting errors in inflation. The goal is ostensibly to figure out how to solve these problems so that forecasts can improve.

• I am by no means an expert in inflation forecasting, and certainly not in climate forecasting. However, looking at say the Mauna Loa carbon dioxide forecasting makes me wonder what impact Covid had on CO2 emissions (eyeballing the chart, it seems close to 0). Even so, the text contains a section stating "In 2020 we also issued an updated forecast once it became clear that the response to the Covid-19 pandemic would cause global CO2 emissions to be much smaller than expected that year." Jun 13 at 20:04
• Yeah, a lot of the climate models will use methods like "data assimilation" to update model predictions with recent data. This is the way that weather forecasters can predict hurricane storm tracks--by adjusting the general climate forecast based upon real-time data from storm spotting aircraft, etc. So I imagine Mauna Loa started to see a divergence between what the general forecast was and what the data assimilation adjusted forecast was, and hence the message. But this shows there are precise ways to measure uncertainty and how that uncertainty affects other measures and predictions. Jun 13 at 20:11
• But my point is that the looking at these forecast values and observed values of CO2, it does look like even a simple ARMA should do reasonable well. What happens to climate models if manmade emissions are NOT the only variable (we talk about roughly 200 years of large increases in C02, whereas most factors like Wilson cycles, Milankovitch cycles, sun activity and so forth did not change much since). If they do, there is great disagreement it seems. Jun 13 at 20:33
• @AKdemy From reading the press release, they seem to say that their prediction in CO2 reduction from covid was a bit too much, because there were still land use changes and cement production. But then the prediction in the rise in CO2 for 2022 was a bit too high, because countervailing forces like El Nino leading to more forest growth--which pulls CO2 out of the air. So the climate models are doing a good job of accounting for the feedback effects from different sources. Climate models do have uncertainty ranges. but those ranges seem to be relatively accurate in this case. Jun 13 at 21:29
• @AKdemy but I don't want to get into a debate about PDEs versus econometric models. The real criterion should be based on model accuracy in predictions. Inflation forecasts must have confidence bounds, but ostensibly, the actual inflation rate overshot those bounds by a probably more than 1.5 standard deviations if not more. Given all the historical data that these models have, that kind of miss seems high. Jun 13 at 21:33

If one could make such long comments, I would write this as a comment, not an answer. I neither forecast weather, nor inflation. However, I am a risk manager in banking, and forecasting is part of my job, at least to a small extent (e.g. modelling non maturity deposits for computing interest rate risk in the banking book and the like). My institution seemed to do reasonable well with NMD modelling, but performed very badly after COVID started. It may be due to my (teams) lack of skills, but rumour has it that essentially all other banks struggled significantly since Covid as well. I believe that modelling people's behaviour works well, until it doesn't. Inflation is essentially a result of interactions of many people in the market place.

Let's look at some CO2 and hurricane forecast data:

In my opinion, this is a textbook example for a simple forecast with very pronounced and regular seasonal swings. It also seems COVID19 had little (to pretty much no) impact on the observed values when looking at the chart only. This pattern seems to be simple to forecast, even ignoring anything related to Covid or El Niño (CO2 seems to rise during El Niño according to this source and all others I checked). Reminds me a lot of Liquor sales forecasting in Elements of Forecasting.

• Hurricane storm tracks are forecasts of their path, only once a hurricane built, and typically last a few days (the storm is mostly gone after a few days anyways). Inflation on the other hand is only computed on a monthly basis and is subject to measurement errors itself.

• Hurricane intensity and seasons are apparently very difficult to forecast. According to a newsletter in 2014 of one of the largest re-insurance companies:

Seasonal forecasts of hurricane activity issued by three prominent forecasting groups have been compared with observations. Looking at the forecast errors since 2001, no significant improvement is visible for any of the examined forecasting schemes.

Forecasts of hurricane landfalls are even more difficult and one of the groups in question (CPC) does not provide them.

hurricane activity exhibits such strong variability that the actual outcome each year often deviates strongly even from the relatively skillful August forecasts.

Despite some unsuccessful years in hurricane forecasting all stakeholders will keep looking at the seasonal forecasts with great interest in the future. The 2013 forecast was one of the worse ever. This could be the result of simply bad luck or the result of important predictors being missed.

MIT's predicting hurricanes seems to also conclude that here are far fewer good options available to predict the intensity of hurricanes because the reasons behind intensity changes are not fully understood and there are many factors involved.

• How would climate models perform if something like the 17th century cold spell (little ice age) would happen again? Since apparently no one knows what caused it (or at least there is no consensus), having the data to forecast another such episode seems daunting. That brings me to inflation since the end of 2021.

Inflation:

• Inflation is a lot more variable and unpredictable than CO2 it seems

• You suggest DSGE models should help forecast inflation. Let's look at the FED's main DSGE model. A stylized description looks like this:

In essence, there is only a handful of Data Series being used in the model.

What was suggested to have caused high inflation?

The ECB believes its a combination of a quick reopening (we did see a large unexpected and unprecedented increase in deposits / savings across the spectrum during lockdown, despite passing on negative interest rates in countries where this was possible, probably because people did not find ways to spend their money), supply chain disruptions (semi conductors for example; which caused record high credit growth rates for inventory financing because "everyone" just wanted to get hold of "whatever they could" to make sure production is not disrupted), and higher energy prices and the Base effect (Inflation is high today because it was so low last year).

I argue, that apart from the Base effect, nothing can be modelled in this DGSE model (unless you assume changes and feed these changes as shocks). While this model may seem extraordinary simple, I invite you to look at the details.

Let's focus on energy prices (not modelled at all in the FED DSGE model for example). The ECB writes that

wholesale gas and electricity prices in particular, annual growth rates in the fourth quarter of 2021 (540% and 390% respectively) were around four times their previous maximum during the period from 2005 to 2020, and all observations since the second quarter of 2021 were well above all previous historical values. Russia’s invasion of Ukraine caused energy commodity prices to increase even further in the first quarter of 2022.

DB Research, prior to the Russian invasion in the Ukraine, states that the reasons for the sharp increases in energy prices are manifold.

• On the one hand, global energy demand has increased drastically due to the synchronised world economic recovery.
• On the other hand, the global energy supply has been severely disturbed by several external shocks and market imbalances such as extreme weather events in the US in early 2021, technical problems at gas fields and pipelines in Europe, low filling levels for gas storage capacity in Germany, interruptions in the global logistics sector or trade issues between China and Australia.

Assuming you could perfectly model demand, modelling the mess of supply shocks is impossible in my opinion. The article also states climate policy measures like the new carbon levy introduced in Germany in 2021, which increased prices for petrol, diesel, heating oil and natural gas by around 5.5%, 7.1%, 16% and 7.7%, respectively (compared the annual average prices for the year 2020).

Now, most forecasters and policy makers believed many of these effects to be transitory. I think it is reasonable to assume no one forecasted that Russia is going to invade the Ukraine, AND forecasted how the rest of the world reacted. However, a good inflation forecast would need to incorporate this major event. For example, it seems the war has a significant impact on car production. That said, the car industry and used cars alike saw major disruptions (PDF download) even prior to the war. Possible reasons are pent-up demand, higher savings rates, shifted consumer preferences due to social distancing making public transport more challenging, and the rise in second-hand online platforms during the pandemic as well as supply chain disruptions (which intensified with the war).

Now given all that info, it is hard to imagine a country does not have much inflation. Yet, Japan still has moderate inflation (2.5% for all items in Apr 2022 according to e-stat.go.jp). It may be difficult to explain why onions increased by almost 100% whereas Welsh onions fell by 16.7% over the year,

but the main question is why Japan, despite being exposed to the same shocks (especially in energy prices, and due to the weakening Yen even more so) as other countries, exhibits almost no pass-through from rising prices to higher consumer prices and wages.

Long story short, what does

Given all the historical data that these models have

help you if none of these causes ever happened before (in tandem)? Or when do you remember the last lockdown, a significant war in Europe that (has to potential to) drastically affect(s) energy and food supply, and major interruptions in the global logistics sector or trade to happen within months?