The May 2011 GDP numbers are out, so it's time for my quarterly attempt to provide an estimate for quarterly GDP growth a month before Statistics Canada releases its first estimates - the most recent exercise is here. This is done by using simple linear regression model of monthly GDP growth on monthly LFS data for employment and hours worked to produce an estimate for (in this case) an estimate for June 2011 GDP. This estimate is then combined with available monthly GDP for April and May to construct an estimate for quarterly GDP for 2011Q2. This isn't quite as *ad hoc* as it sounds; Statistics Canada recently published a note exploring the informational content in the GDP numbers of different frequencies (and check out footnote 2).

But it turns out that the point estimates themselves aren't that bad, at least, not compared to Statistics Canada's own preliminary estimates:

Quarter WCI estimate First StatsCan

releaseLatest data 2009Q1 -6.9% -5.4% -7.9% 2009Q2 -3.4% -3.4% -3.7% 2009Q3 -0.4% 0.4% 1.7% 2009Q4 4.0% 5.0% 5.0% 2010Q1 5.5% 6.1% 5.6% 2010Q2 2.7% 2.0% 2.3% 2010Q3 1.5% 1.0% 2.5% 2010Q4 1.9% 3.3% 3.1% 2011Q1 3.8% 3.9% 3.9% 2011Q2 0.1%

Even if you include 2011Q1 - where the most recent data *is* the first StatsCan release - it turns out that the WCI estimate is a slightly *better* predictor of the most recent quarterly GDP growth data - in the RMSE sense - than is the preliminary StatsCan estimate.

So I'm going to keep on doing this. The WCI estimate for annualised GDP growth in 2011Q2 is **0.1%**. This is quite a bit lower than the Bank of Canada's estimate of 1.5%.

Coo! Footnote 2. A genuine StatsCan cite. Well done!

But your latest estimate is depressing.

Posted by: Nick Rowe | July 29, 2011 at 10:43 PM

Have you ever decomposed your forecast error into (i) your error in forecasting the last month's GDP using your model, (ii) the revision of the first two month's estimate of GDP, (iii) the difference between the three monthly estimates of GDP and the quarterly value?

Posted by: Angelo Melino | July 31, 2011 at 07:16 AM

No, but I will from now on. I've been replacing the old data with revised data, so I can't reproduce my old numbers. I'll start archiving my programs as I go along so that I can do that sort of breakdown.

Wish I had thought to archive them before.

Posted by: Stephen Gordon | July 31, 2011 at 08:13 AM

The Statscan article highlights the importance of using two months of the first quarter (February and March; ignore January's growth) and as many of the three months of the second quarter as are available in order to calculate what is currently known about the second quarter. By doing this, analysts can take advantage of the statistical relationship that exists between monthly and quarterly growth rates. (Although as noted in an earlier post by Stephen Gordon, monthly and quarterly GDP are based on different concepts.)

By using the -0.1% of February 2011, the 0.3% of March 2011, the 0% of April 2011 and the -0.3% of May 2011, the formula (included in the article) indicates that you are assuming a 0.2% increase (approx) in June 2011 in order to have a 0.1% annualized GDP in the second quarter.

April would normally have the largest weight of all of these months in determining the second quarter's growth rate but there was no change in that month, and March and May, which have similar weights in determining the second quarter's growth rate more or less offset each other. Therefore, the difference between June's growth rate and that of February's -0.1% (these two months again have similar weights) will define the quarter's growth.

(Also, the old data can be found in the tables attached in the Daily archives for monthly and quarterly GDP).

Posted by: Diana Wyman | August 17, 2011 at 05:20 PM