Logout / Access Other products Drop Down Arrow
Get live help Monday-Friday from 7:00AM-6:00PM ET (11:00AM-10:00PM GMT)  •  Contact Us
Check out our new FAQ section!
RSS Feed
Using Data Buffet: Caveats when converting low frequencies
Wednesday, 27 Jan 2016 19:11 ET
By Phillip Thorne
Data Buffet enables you to convert a time series with a "low" frequency (e.g., annual or quarterly) to a "high" frequency (e.g., monthly). This does not select or create better data; rather, it is a mathematical interpolation, and must be used with caution, as follows.

This document explains the tools in Data Buffet that change the display of a time series from its native frequency to a synthetic frequency that may be useful for analytic or comparative purposes.

What's a native frequency?

Every time series in Data Buffet, both historical and forecasted, has a particular frequency at which the data is observed: each day, week, month, year, etc. Some important economic indicators are reported at a low frequency, especially for smaller areas. For example, most Moody's Analytics forecasts have a quarterly native frequency [1].

What happens if I change the frequency?

If you use the "Frequency" drop-down in View or Basket modes, you're either condensing data points (a high-to-low frequency conversion) or interpolating new ones (low-to-high).

What does frequency conversion not do?

  • Create new data with economic expertise.
  • Produce estimates for intermediate periods [2].
  • Retrieve alternate time series (actuals) with a different native frequency [3].
  • When converting forecasted series, retrieve corresponding historical or estimated series.

How can I check the native frequency?

There are four ways:

  • Examine the last letter in the concept code (the part of a mnemonic before the period and geo code). Annual/quarterly/monthly/weekly/daily series are denoted by the letter A/Q/M/W/D, respectively [4].
  • Use Mnemonic 411.
  • Open in View mode and examine the "Frequency" dropdown.
  • Include the "Frequency" field in a basket.

How do low-to-high and high-to-low frequency conversions work?

Time series can be converted in two directions:

1. Up-conversion: Low-frequency data (annual, quarterly) to a higher frequency (monthly). This interpolates the data by computing a curve through the actual observations and picking new points off the curve [5]. This curve may exhibit implausible properties, especially at the ends [6].

2. Down-conversion: High-frequency data (daily, weekly, monthly) to a lower frequency (quarterly, annual). This reduces the number of data points, usually by taking an average or sum (depending on the series) [7].

In either case, the exact nature of the curve is controlled by the Observed property of the time series.  For up-conversion, the choice of Conversion Method in Basket Options is also significant.

The process of frequency conversion is necessarily generic, so do not expect the converted points to exactly match any related time series.  An ostensibly comparable indicator from a third party may or may not match Data Buffet's output, depending on specifics.

Do alternate frequency series exist?

Some indicators are available in families with several native frequencies, but you have to manually select them. Such families usually consist of a fundamental high-frequency indicator, plus lower frequencies for convenience. For example, the FRB H.15 "Selected Interest Rates" release is available at business-daily, weekly and monthly frequencies (IRFEDD.IUSA, IRFEDW.IUSA, IRFEDM.IUSA), and the latter two are averages provided by the source. Similarly, mortgage yields are published weekly by Fannie Mae, and Moody's Analytics precomputes a monthly version.

Also, Moody's Analytics produces estimates for certain major indicators, such as CPI, housing stock, and retail sales. We do so when the frequency, geographic coverage, or industry detail of the source data are inadequate.

How does this affect forecasts?

Most of our forecast models are quarterly, and each forecast time series may be based on one or more historical or estimated series (drivers) of any frequency.

If the forecast series is a direct counterpart to a quarterly historical time series, the actual values are included in the forecast. Use the Historical end dates Basket Option to see where the switchover is. (Not available for all forecasts.)

If the forecast is based on a single historical time series that's monthly, you can't directly compare the two. The frequency down-conversion that creates the forecast has destroyed some of the input data, and it cannot be reconstituted. If the driver is erratic, the forecast will be smooth.

See also


Note 1, Update vs. native frequency. The forecast models are updated ("run" or "spun") monthly with the newest driver values, but this is distinct from the native frequency of the series.

Note 2, Estimates. Moody's Analytics produces estimated time series for many indicators when the source data is sparse in time, space, or industrial detail. These are typically identified by an "R" prefix on the concept code. The estimates often serve as the driver for forecasted series; and sometimes the estimate is monthly and the forecast is quarterly. Conversion will not retrieve the higher-frequency series.

Note 3, Alternative frequencies. Some indicators in Data Buffet are available in families with multiple native frequencies. These alternatives are never retrieved by the Frequency Conversion controls.

Note 4, Concept codes lacking AQMWD. The concept codes for certain popular U.S. indicators don't have frequency suffixes. For example, employment level from the BLS CES is ET, and is monthly. The concept codes for our international data, however, are completely standardized in this respect.

Note 5, Cubic splines and alternates. By default, low-to-high frequency interpolation is performed using a geometric construct called a cubic spline, which usually produces the best results. Alternative methods are sometimes appropriate and are available by using the File»Conversion Method Basket Option, or by the Basket function (convert), but their use is recommended only for advanced users. For details, please contact us.

Note 6, Nonsense curves. In up-conversion, when the actual points in a time series are trending downwards, a curve drawn between them may dive into negative territory -- even if that doesn't make physical sense. This may happen at the end of the time series, or sometimes in the middle.

Note 7, The Oobserved attribute. High-to-low conversion is performed according to the Oobserved attribute of the series (average, sum, end, etc.), a property stored in our database. The attribute is visible only in a basket, and only if you select the optional field Conversion Method. It may be overridden using the (convert) Basket function.

The Observed attribute depends on the nature of the indicator. For example, flow observations are combined as an average; per-period levels are combined as a sum, but individual year-to-date levels pick the end-point.