******************************************************************************
                    Summary of clock prediction strategy
                                   Jim Ray
                         22 May 2001 12:25:52 EDT

                  (update of IGS Mail #2962, 31 Jul 2000)
******************************************************************************

The procedure used for our clock predictions is to extrapolate the
estimated satellite clock values from the observational data preceding
the prediction period.  The observation period is nominally the 2 days
immediately before the predictions.  The 1st day before the predictions
is alway used, and an earlier 24-hr solution using the same reference
clock is sought.  The search priority for a matching solution is: 1 day
earlier, 1.5 day earlier, then 0.5 day earlier.  If no matching solution
is found after those attempts, then only the 1-day solution before the
predictions is used.

Using all the data for each satellite clock individually, fits are made
using the following models:

      linear + sinusoid              for satellites with Cs & IIR Rb clocks
      quadratic + sinusoid           for satellites with II/IIA Rb clocks

where the period of the sinuoid equals the orbital period (~12 hr).
Some satellite clocks show very pronounced sinusoidal variations (most
prominently PRN06) while most show little.  The satellite clock types
are maintained in a master file which is checked at runtime.

Continuity between the observed clocks and the predictions is enforced at
the last epoch of the observation period.  If the RMS of the fit to the
observed clock of any satellite exceeds 10 ns (nominally over 2 days),
then no clock predictions are reported for that satellite.

The USNO clocks are estimated relative to the clock state of a chosen
tracking receiver (equipped with a H-maser external standard), which is
not adjusted as the reference.  Note that this reference procedure can
cause occasional problems for both our observed and predicted clocks,
for example when the chosen reference clock suffers a reset.  In this
case, no clock predictions will be issued.  For this reason, it is
vital for several analysis center to submit clocks for the IGU
combination using different clock references.

Future prospects -- Errors in IGU clock predictions grow roughly linearly
with time (rather than as sqrt(t) expected for a random walk process)
because the dominant errors are due to misfits of the clock model to the
observed clocks and to jumps in the satellite frequencies.  So it should
be possible to improve the IGU clocks by shortening the update cycle.
However, for the highest accuracies it will be necessary to observe and
broadcast the satellite clocks in a near real-time mode.