Nowcast bleg

I just received in my email box the latest “Nowcast” from the New York Fed. Unfortunately, I cannot understand what they are trying to say. It seems like they are predicting (as of July 31, 2020) that Q2 growth will be minus 13.75%, at an annual rate. But one day earlier, growth was officially estimated to be minus 32.9%. So what is this Nowcast telling me? Is this an estimate of how the GDP number will eventually be revised?

After receiving suggestions, I may add some further comments below. But I don’t want to waste time if I’m misinterpreting the Nowcast:



12 Responses to “Nowcast bleg”

  1. Gravatar of Garrett Garrett
    31. July 2020 at 11:28

    For 1Q20 and 4Q19 the charts also include latest revisions, so I agree that it could be interpreted as what they believe the final revision will be.

    But it’s probably just that they have a team whose job it is to run this statistical model and post the results without comment. Over time the “real” forecasters can use the final error to get information on how predictive the underlying features are, which could help them improve their “real” models.

  2. Gravatar of Christian List Christian List
    31. July 2020 at 11:39

    I have read an article in German, don’t know where it was exactly right now, but the article said that
    the minus of 32.9% in the US would be calculated differently in Germany.

    Translated into the German system the US minus would be more in the range of 10-13%, which would then also be equal to the German minus right now.

    One point of the article was that the German minus and the American minus are currently pretty much the same, but most media outlets would report it wrongly because of the different systems in use.

    Maybe Nowcast uses this different German/European calculation method as well?

  3. Gravatar of John Hall John Hall
    31. July 2020 at 11:42

    The NY Fed Nowcast releases a new forecast every week using the data available by Friday. They have been forecasting around -15% or so for several weeks. You are not misinterpreting anything. They were just way off. This is partially due to the methodology used.

    The NY Fed Nowcast uses a multi-factor Kalman filter approach. If I recall correctly, GDP is just one additional variable as part of the model. The nice thing about the NY Fed Nowcast is that it provides a quarter ahead estimate quite early.

    By contrast, the Atlanta GDPNow did a much better job at capturing what was happening this quarter than the NY Fed Nowcast. The Atlanta GDPNow uses bridge equations to build up higher frequency data into estimates of the quarterly components of GDP and then calculates GDP as GDP itself is calculated.

    Now there a subtle thing that you were also asking about, which is whether they are trying to forecast the advance estimate or the final estimate. I don’t recall this specific detail for each of these models off the top of my head, but I would argue that the reason it is off by so much would not have anything to do with this difference. The potential revision here is likely to be in the range of a few percentage points, not 20%. For comparison, Goldman Sachs was forecasting around -29% for the advance estimate and around -33% for the final estimate. Not every Wall Street economist though provides the breakout.

    I would also note that Wall Street economist forecasts were all close to the Atlanta GDPNow estimate (and hence the actual result) and none were close to the NY Fed one.

  4. Gravatar of John Hall John Hall
    31. July 2020 at 11:44

    @Christian List That is not the issue here. Both report in QoQ annualized. Americans will often convert European GDP QoQ numbers to QoQ annualized, and vice-versa.

  5. Gravatar of Ray Lopez Ray Lopez
    31. July 2020 at 12:16

    SSumner: “Unfortunately, I cannot understand what they are trying to say”… seems to be a reoccurring theme.

    I think Christian List is right and John Hall is wrong. Crudely, -13.75% in one quarter = -4×13.75 = -55% for the year.

    Anyway the obsession with historical GDP in a time like this is absurd. The relevant inquiry is what will happen going forward.

  6. Gravatar of Steve Steve
    31. July 2020 at 12:28

    Hi, I can help you out. I have implemented several different nowcasts including one very similar to the NYFed model.

    The issue with forecasting 2020 Q2 GDP is that monthly GDP has been historically volatile. April GDP was down almost 10 percent month/month, whereas May GDP rebounded about 5 percent month/month and June GDP increased a smaller amount.

    We cannot take the geometric average of the monthly GDPs to get a correct q/q GDP. A correct estimate of GDP would have to use the Q1 level, then nowcast a Q2 level and from these determine the annualized growth. This is not how the NYFed nowcast works. Everything in the model is based on growth rates and not levels, so it loses the levels needed in the current situation. This is akin to using a linear model when linearity clearly doesn’t hold.

    The NYFed GDP nowcast is implicitly giving more weight to the improvements at the end of the quarter relative to the historically bad data received in the first month of the quarter. Structurally the current -14% print reflects all the information available up through now (except for the actual Q2 GDP release), but is hampered by limitations in the methodology.

  7. Gravatar of ssumner ssumner
    31. July 2020 at 19:50

    Thanks Steve, Doesn’t sound like a very good model.

  8. Gravatar of Nathan Nathan
    31. July 2020 at 23:57

    Scott, since you don’t want to read the article, I’ll give you the answer right here.

    For any readers who don’t want to be spoiled, you should read the article, including the additional material linked at the end of the article.
    For those who don’t mind spoilers, keep reading.

    The prime age labor force participation rate is about 4-5% lower in the U.S compared to peer countries. There are two factors responsible for that gap.
    The first factor, which accounts for about two of those percentage points, is due to the higher effort requirements firms are imposing on their employees. Making people work harder than they actually want to, naturally robs work of some of its utility, and leads those who don’t need to work to drop out of work and rely on their spouse for income.

    The second factor, which accounts for about 2.5% of that lfpr gap, is due to the far greater competition for jobs than normal (we’re still talking pre-pandemic, so 2019). This forces people to apply for more jobs and attend more interviews to get just one job offer. The unemployment rate in 2019 was 3.5%, but for the job seeker it would be as difficult to get a job as during the recessions of 2001 and 1990. I.e when unemployment rose by 2.5%.

    So that is your answer Scott.

  9. Gravatar of Steve Steve
    1. August 2020 at 04:58

    Hi Scott,

    I would say under normal circumstances it is a good model. Technically it is very impressive and it solves the problem of filling in missing observations using the Kalman Filter.

    But, these are not normal circumstances. The implicit assumptions underlying the model no longer hold.

    To be successful now, the model would have to forecast monthly GDP. That is a more difficult problem, because the best data is also at a monthly frequency, with lags. There is not enough higher frequency data with long enough histories to build a monthly model comparable to the quarterly one.

  10. Gravatar of Michael Sandifer Michael Sandifer
    1. August 2020 at 06:06

    Yes, at times like this, market forecasts are even more important, though real shocks complicate things. Nominal shocks are extremely easy to deal with.

  11. Gravatar of Michael Rulle Michael Rulle
    1. August 2020 at 07:03

    Maybe Trump is playing 3D chess. NPR reports that 50% of people under 35 do not know how or have the resources to mail a vote. LOL x 3

  12. Gravatar of ssumner ssumner
    1. August 2020 at 09:20

    Nathan, I read the whole article, I just didn’t find it persuasive.

    Steve, I defer to you on that point, as I have not studied the issue closely. But keep in mind that models will be judged mostly how they do in extreme cases. No one cases about Nowcasts in normal periods, the whole point is to get an early read on major shocks. That may be unfair, but that’s how people make judgments.

    Michael Rulle. Apathy among the young got him elected in 2016, maybe they’ll do so again.

Leave a Reply