Dear all,

I followed up on the suggestions at the conference call as follows:-

1) USE LESS DATA

Unfortunately, kicking out only 1990 makes the overall equation a lot less 
robust, in fact dramatically so, and so eliminates the possibility of using 
less data.  The model tested was the RPI(month+15) & Deviations from 
long-term average for Brent Crude.  The r-squared and F-statistic collapsed 
by eliminating the first 12 months of data.

2) DEVIATIONS IN CRUDE VARIABLE

Shifting the crude explanatory variable backwards and forward by 3 and 6 
months did not alter the model goodness-of-fit parameters dramatically and so 
my suggestion is that we stick with the following model:

PLLU[t] = a.RPI[t] + b.RPI[t+15] + c.BrentCrudeDeviations[t] + Constant

3) TESTING MODEL WITH SINE WAVE

To gauge the response of the model to wildly-varying RPI forward curve, a 
sine wave of period 3 years for RPI was input into the PPI model.  The result 
was as expected; PPI pre-empts the moves in RPI by about 8 months.  The 
magnitude of the oscillations is also reduced.  This shows that if we had 
more detail in our RPI forward curve, then the PPI model would reflect those 
peaks and humps adequately.



CONCLUSION

I therefore propose that we use the model that incorporates RPI, RPI[t+15] 
and deviations of Brent Crude from long-term average.  The new model is 
plotted below in burgundy and can be compared to the old PPI which is 
depicted in blue. 

The new model achieves the two main objectives of the PPI curve: it is 
significantly more robust and stable than the existing one, and it is 
considerably less sensitive to the input coefficients, which results in us 
having more confidence in our monthly P&L.


Please note that all this analysis only applies to PLLU, and that a separate 
study will be needed for the DZCV PPI index.

Regards,

Anjam
x35383

---------------------- Forwarded by Anjam Ahmad/LON/ECT on 09/03/2000 16:52 
---------------------------


Anjam Ahmad
08/03/2000 14:03
To: Martina Angelova/LON/ECT@ECT, Harry Arora/HOU/ECT@ECT, Maureen 
Raymond/HOU/ECT@ECT, Zimin Lu/HOU/ECT@ECT, Farouk Lalji/HOU/ECT@ECT
cc: Trena McFarland/LON/ECT@ECT, Dale Surbey/LON/ECT@ECT, Stinson 
Gibner/HOU/ECT@ECT, Vince J Kaminski/HOU/ECT@ECT, Leandro 
Ibasco/Corp/Enron@Enron 

Subject: Update on PPI Model for Inflation Book

Dear all,

We thought it might be useful to incorporate Brent crude as an explanatory 
variable for PPI; it was found that deviations of Dated Brent Crude from the 
long-term average of $18.80 was the best form of the variable to use (for 
predictions the Brent forward curve produced by the Global Products Group is 
used).  The three new equations developed were:-

PLLU(t) = a.RPI(t) + b.RPI(t+N) + c.(DatedBrentCrude - 18.8) + constant, 
where N is 14,15 or 16
[REDDISH CURVES]
r-squared approx 0.49
F-stat approx 32

The chart below shows what our projected PLLU curve would be given this 
equation, and also the three best relations from before which were based upon 
current and future RPI:

PLLU(t) = a.RPI(t) + b.RPI(t+N) + constant, where N is 14,15 or 16
[GREENISH CURVES]
r-squared approx 0.47
F-stat approx 45

COMPARISON OF MODELS
As you can see, the two equations differ in the very short-term and very 
long-term; the inclusion of deviations of Brent Crude leads to short-term 
predictions of 3.0% to 3.2% over the next six months.  The greenish curves 
predict PLLU in the range of 2.5% to 2.8% over the next six months.

The curves are then very similar until 2009, when the models including crude 
break-away to the upside, relative to the falling RPI curve.  The model based 
purely on RPI hugs the RPI curve much more closely in the longer term.  This 
is only important to the extent that we have large positions beyond 2009 
(which we don't).

SUGGESTION
What could be useful now is a differently-specified model designed to 
forecast only the next 3 months, using auto-regressive or auto-regressive 
error terms.  This model would be far more accurate in the near-term, and we 
could include this information onto the front of this long-term model.  This 
may be useful, despite the fact that most of our exposure is in future time 
buckets.



BACK-TESTING

All the models give similar visual and statistical performance over the data 
sample used (based mainly on 1990s "new paradigm" economy).



Hopefully we can discuss these and other points later in the tele-conference; 
your ideas on this would be appreciated.

Regards,

Anjam
x35383