I've downloaded the corresponding forecast concept from several alternative scenarios and their histories (actuals) don't match. Why not?
The short answer
1. Most likely, you converted the set of series away from its native frequency, e.g., quarterly [1] to monthly. Given the mathematics of frequency conversion, a difference in one part (i.e., the forecast segment of each scenario) will affect the entire series.
2. Or, you are comparing different vintages.
3. Otherwise, please contact us.
The long answer
Each Data Buffet forecast time series consists of a historical segment and a forecast segment. The historical segment is populated from the “actual” data [2], is fixed [3], and is identical across all scenarios—but only at the native frequency. Once you perform a frequency conversion, the historical portion is no longer inviolate.
Low-to-high frequency conversion is an interpolation (i.e., it creates new values between the original periods) using a technique called a "cubic spline." For each targeted period, the algorithm examines several adjacent periods—approximately seven to each side. When converted, these adjacent intervals are affected. Therefore, changes to values in the forecast segment will propagate backwards to the historical segment—around 20 months prior to the boundary [4][5].
This is mathematically necessary to create a smooth curve crossing the historical/forecast boundary (i.e., the conversion cannot be piecewise). If the future is different, it will affect the past. To see the undistorted historical values, you must use the native quarterly frequency.
1. Most Moody's Analytics forecast series have a native frequency of quarterly.
2. "Actual" data in the historical segment: A forecast series will have one or more drivers which are third party-historical or Moody's Analytics-estimated time series. The historical segment is populated entirely from, but is not necessarily equal to, any particular driver (see related article).
3. The historical segment is immutable insofar as the drivers are; if the drivers are revised, so too will be our forecast series. For example, home price indexes using the "repeat sales method" typically restate their entire history with each update.
4. Technically, the divergence will cover the entire the historical segment, but to a minute degree that is either (a) statistically insignificant or (b) obscured by your decimal setting.
5. The "last historical point" is listed in Mnemonic 411.