CRSP Index Percentages as of 8/31/2026

 

Using 12 Mos. Factor table vs. Historical REFCPI-based ratios.

September 4, 2026 provided a nifty natural experiment on how the market prices seasonal adjustments for dates in the near future. In particular near dates beyond the max REFPI date (1-OCT-2026).

I traditionally used a 12 month table with a day 1 value and daily increment based distance to the next month (1-SEP to 1-OCT). This used either historical averages or the last 12 mos. ending on the max REFCPI date. Thus in in the 12mos. model, for today's case, the 1-OCT value is based upon the most recent July 2026 CPI report. The daily increments in OCT are calculated against 1-NOV-25 REFCPI. Giving a daily increment of -0.0000387.

KevinM's Treasury Portal uses historical values only, not a 12 mos. table. For 11 months a year, the recent table approach and the historical approach produce identical results. Except for the month of October (at this test point). The reason is that the table tool calculates the daily delta based on the value for 1-OCT-26, and 1-NOV-2025 and historical approach calculates them based upon the daily delta between 1-OCT-25 and 1-NOV-25. The daily increment for this historical model would be -0.0000197 or about half of the amount.

For context, a larger decrement will reduce the price more, and thus produce a higher adjusted yield. In months with positive daily changes yield impact will be reversed.

This results in meaningfully different SA Factors for 15-OCT.

Normally, these values are not that different. This year the deltas between OCT-NOV for 2025 were quite a bit different than for the historical averages, coming in at half the daily increment. The table model is the highest increment for any of my historical averages, and the 2025 REFCPI value is the smallest of any of my historical averages.

If the curve is fitted against the historical REFCPI model better than the monthly Factor model, it is a good case that the market is more closely pricing to the history model, not a table model.

If both were off the curve significantly, that may imply a better factor would to use an average. Since the shorter-term averages are closer to the REFCPI value, more back-tests may be necessary to completely rule out historical averages.

And below is what happened yesterday (I have seen this hook on previous days as well, but did not have the tools to investigate). It appears that the historical value matches better. Now, this is just one date, but it does provide a reasonable explanation for some well known misbehavior in at least I have seen in very short-term maturities. I am going to continue to watch this a bit more before deciding it is completely settled. 

Note there are visibly better matches to the yield curve on both the 15-OCT-26, and 15-OCT-27 points on the curve.