In my last post I claimed consumption spending is getting back to normal. After a volatile 2020 and early 2021, consumer spending appears to be falling back to long run trends (i.e., trends that prevailed before Covid-19).
On the heels of that post, the BEA released some important data. First, yesterday they put out their “advanced estimate” of real GDP for the second quarter. That value is -0.9 percent. That is not great, but it is not terrible either considering the level of macroeconomic angst that has gripped the country the past few months. (Recall, in the first quarter real GDP fell by 1.6.) The GDP report showed, too, that real personal consumption expenditures for the second quarter increased by 1.0 percent.
The second important piece of information came out this morning—the June estimate for consumer spending. In June, total real consumer expenditure increased by a tepid 0.1 percent. On the bright side, this was rebound from the decline in consumption in May.1
The June estimate, along with the quarterly increase of 1.0 percent for consumption expenditures contained within the GDP report, suggest that consumers are holding steady. In spite of a quarterly decline in spending on goods, consumers are at least still spending on services (see Table 1 in the GDP report).
As I have explained before, consumer spending is a key chunk of GDP growth, so understanding what consumption is doing, and will do, is crucial to guesstimating the future path of our macroeconomy.2
In this post, I use the updated consumer spending data to forecast what “normal” might look like for the next few months of 2022. Below, I’ll provide the forecasts, explain them, and then at the bottom of the post I will provide details on how I created the forecasts (for those that are interested).
The Long(er) tail
Here’s the thing about consumer spending: it’s very persistent. That means we as consumers generally maintain a pretty consistent amount of spending from month to month. This is especially the case for nondurables but it is also apparent when you look at the total (which is comprised of nondurables, durables and services). For example, check out the total value of consumer spending for each month over the past year (from July 2021 through June 2022):
The total amount of consumer spending each month has hovered between about 13.6 trillion dollars to just under 14 trillion dollars. Keep in mind, too, these values represent new spending each month—so there was 13.912 trillion (constant) dollars of spending alone in June of 2022.
The changes from month-to-month of those trillions of dollars are what I focused on in my last post, and which we can see here updated with the value for June of 2022:3
Why are the past 12 months particularly important? Because I am going to base my first forecast of consumer spending on the growth rates for each of those months. In macroeconomics a typical forecasting model will rely on past values of a variable to forecast its future. I discussed this a little bit at the end of this post, but for a quick picture of a forecasting model, you can think of it like this:
The “weighted average” means that some past values matter more than the others—in theory the most recent values matter more than values from farther back in time. Using the past 12 months of data as the inputs in my forecasting model, my monthly forecasts from July to November are as follows:
Notice that for July, I predict an increase in consumer spending of only 0.03 percent, so essentially zero growth (recall, we are forecasting for July since that data will not be released until the end of August). For August, the model predicts 0.22, and so on out to November. You will notice that the pattern of the forecast from month to month looks similar to the most recent months, from February through June. That is not surprising since as I mentioned above, in a forecasting model like this the most recent months are likely to matter the most. Overall, the average month-to-month growth for the next five months is about equal to the previous 12 months.
The Short(er) tail
Why did I choose the past 12 months as the inputs? In many macroeconomic studies, past values of a year or more are often included in forecasting models. So I started with a year’s worth of inputs. However, research on forecasting consumption is not conclusive about that, so I also considered a shorter time period. The following forecast was created with only the past three months included as inputs:
In contrast to the forecast using twelve months of past data, here the forecast is “smoother” over the five-month forecast horizon (July through November). The forecast average of 0.17 from the “shorter” model is a bit higher than the average of my first forecast (0.13).
Okay, so what?
What can you take away from this forecasting exercise? Here are few thoughts.
1. The forecast using the previous 12 months as inputs has a pattern similar to the history of data. The forecast for the three-month model is smoother, appearing to “split” the difference of the April to June growth rates. What do these patterns reflect? Simply put, the patterns reflect the history of the chosen inputs, as well as how I estimated the weights in the model (see the postscript to this post for those details). In general you must understand this about forecasting models: even though we are creating a forecast, in truth, most forecasting models are backward-looking in a mathematical sense.
2. The most “reliable” estimate is the one-month ahead forecast. In this case, the estimate for July is the most reliable since it is estimated from actual data (the values for June, May, and so on). The next monthly forecasts—two months ahead, three months ahead, etc.—are more problematic since those forecasts contain forecasted values as inputs. For example, the most recent month input for forecasting the month of August is the value for July. But, the July number is itself a forecasted value. In my first forecast, by the time you estimate the forecast for November, there are four forecasted inputs in the model. For the second model with only three months used as inputs, all of the inputs used for the November estimate are forecasted values. That doesn’t bode well for the reliability of the November forecast.
3. What is a soothsayer to do? To overcome the deficiency discussed in points 1 and 2, you can conduct “scenario” analysis. For example, say you want to estimate the value for November. Instead of using forecasted values for the months July through October (as I have done above), you could simply say something like, “assume that consumer spending falls 0.25 each month from July to October.” You plug those values into the model and get your November forecast. Call that your “pessimistic” forecast. Then you could say, “Okay, now assume that consumer spending increases by 0.25 each month from July to October.” You plug those values in to generate a more optimistic November forecast and then compare that to your “pessimistic” forecast. In this way, your forecasting estimates need not rely solely on the history of the series, which can help you or your business (say) plan for alternative scenarios.
4. For the reasons just stated, forecasting can be a frustrating endeavor. The forecasts above are predicated on the assumption of “holding future shocks constant.” But we know there will be future “shocks” or things that will either rattle consumers or boost their spirts. We know that will happen. We are not total fools, right? But we do not know exactly what will happen, when it will happen, nor how consumers will respond to whatever that thing is.
5. You might be wondering—especially in light of the Fed’s aggression this past week—“Hey, know-it-all, what about rising interest rates? Shouldn’t that be in your model, dummy?” That is a great question, albeit a little harshly worded and needlessly personal. Yes, that is on my mind. I plan on addressing that in follow-on posts to this one. I also still need to consider how consumer sentiment may or may not help forecast consumption.
I am saving those topics for later since this post is already too long. Writing about forecasting and having it be too long is a double-whammy kiss-of-death for any hopes that people will read this post or my blog. It is not a great strategy, but I can’t seem to help myself. So, if you actually have read to the end of this post, thank you.
Forecasting Assumptions
Here are some details on what I did to create the forecasts in this post.
1. To generate the forecast I first estimated the consumption model over the 2002 to 2019 period. I consider that era the baseline or “normal” for consumption spending. As I noted in my last post, the year 2020 and the early part of 2021 are anomalies.
2. By “consumption model,” I mean that I estimated with consumption as the dependent variable, and twelve-monthly lags of that variable as the independent variables. In forecasting, you call that an “AR(12)” model or an “autoregressive model with 12 lags.” For my “short(er) tail” forecast, I estimated an AR(3).
3. For the forecast, the future value of monthly consumption is a weighted average of the past month values. So, a forecasting equation for next month’s consumption would look like this:
. . . and so on, for as many months in the past that you want to go. The fancy looking w’s are the weights on each month.
4. The values for the weights came from the estimated regression model mentioned in point #2.
5. The values for “June’s value,” “May’s value,” and so on came from the actual data (values which can be seen in the figures above).
Remember, these values are real numbers with prices held constant, meaning consumers purchased 0.1 percent more stuff from June 1 through June 30.
A third very important piece of macroeconomic information also came out this week: the Fed juicing interest rates another 75 basis points. On top of everything else, all of this macro news nearly overloaded my synapses. So, I am saving discussion on the Fed, interest rates, and consumer spending for next week.
As part of the June report, the BEA updated their estimates for April and May. In my last post, the graph showed May with a -0.4 decline; the update now shows May declined by 0.3.
Thanks for the read -- very interesting!