Mystery Of July Retail Sales “Beat” Solved: It Is All In The “Seasonal Adjustment”
The July retail sales beat came as a surprise to many: an 0.8% increase (full series here) at a time when the data was supposed to grow at less than half this would surely be indicative of a potential turnaround in the US economy. Then we decided to do a quick spot check if maybe the Census Bureau had not adopted one of the BLS’ worst habits: fudging seasonal adjustment factors. The reason for this is because we happened to notice that Not Seasonally Adjusted (full series here) retail sales data in July actually declined by 0.9% from $405.8 to $402 billion. Of course, if the Census Bureau was using a consistent, or at least remotely comparable July seasonal adjustment factor as it has in the past, this would make sense and we would move on. So we decided to look at what the July seasonal adjustment variance over the past decade has been. What we found would have shocked us if indeed this is not precisely what we expected:with the July seasonal adjustment factor routinelysubtracting a substantial amount from the NSA number, averaging at -$5.2 billion, in 2012, for the first time this decade, the seasonal adjustment not only did not subtract, but in fact added “value” to the NSA number, resulting in a seasonally adjusted number that was $1.9 billion higher than the NSA number at $403.9 billion.
So what would have happened if instead of arbitrarily deciding to add a seasonal contribution for the first time in a decade, the Census Bureau had used the last decade average factor of $5.2 billion (not adjusted for inflation, so the end number would be far greater)? Instead of rising by 0.8% Seasonally Adjusted retail sales would have declined from $400.7 billion to $395.5 billion, or a 1.3% decline.
And that is how data is fudged.
Those curious what the model behind this now glaringly obvious seasonal adjustment fudge is, read on (source):
We use the X-12 ARIMA program to derive the factors for adjusting data for seasonal variations and, in the case of sales, for trading-day and holiday differences.
Adjustment of estimates is an approximation based on current and past experiences. Therefore the adjustments could become less precise if current competitive pressures, changes in consumer buying patterns during holiday periods, and other elements introduce significant changes in seasonal, trading-day and holiday patterns.
Each month for sales, concurrent seasonal adjustment uses all available unadjusted estimates (including the latest preliminary and advance estimates) as input to the X-12 ARIMA program. Factors derived from concurrent seasonal adjustment for sales are applied to the unadjusted advance, (one month after the preliminary) preliminary, and final (one month before the preliminary) estimates and to the previous year estimates that correspond to the advance and preliminary months.
The table Combined Seasonal, Trading-Day, and Holiday Adjustment Factors for Retail and Food Services Sales by Kinds of Business presents the combined seasonal, trading-day, and holiday adjustment factors that are used to adjust sales estimates. For kinds of business whose last observation is an advance estimate, two months of projected factors are shown. For all other kinds of business, three months of projected factors are shown. Projected factors are estimates of the factors that will be used to derive adjusted estimates when unadjusted estimates become available.
More information about the X-12 ARIMA Program is available on the Census web site athttp://www.census.gov/srd/www/x12a/ .