Comparative performance of two valuation methods (per TBA pass-through)

Task/Operation	Quasi Monte-Carlo *)	Backward Induction (APD)
OAS accuracy (pass-throughs)	1-2 bp with 2000 paths	Exact; practically < 0.5 bp **)
Computing OAS given Price	0.8 sec	0.1 sec
Computing Price given OAS	0.85 sec	0.3 sec
Computing Duration, Convexity	Two more sets of runs	No additional time
Computing shocked prices	New set for each scenario	No additional time
Standard full-scale analysis ***)	28 sec	4 sec
Above, without KRDs and Vega	14 sec	0.8 sec

*) Quasi Monte-Carlo is proven to be twice more accurate than standard Monte-Carlo, for the same number of paths
**) Small error is caused by prepay lag, a residual source of path-dependence
***) Includes OAS, Duration, Convexity, 5 Key Rate Durations (KRD), Vega, and 5 Prepay Tuning Durations as reported weekly on the AD&Co website.

The relative performance of backward induction versus Monte-Carlo cannot be measured by a single number. For example, our standard TBA analysis includes computations of Vega and KRDs that require multiple rebuildings of the interest rate grid. This operation is common for all valuation methods and represents a “fixed cost” built-in the running time. At no surprise, omission of Vega and KRDs benefits backward induction more than Monte-Carlo.

Having appreciated the speed and accuracy of our APD model and its backward induction implementation, one should keep in mind that AD&Co offers a fully adequate Monte-Carlo and even better quasi Monte-Carlo for structured deals (CMOs, ABS). These techniques have not changed from v5.2, so I will skip their descriptions and features for the sake of brevity.

OAS is Good, but prOAS is Better

Doesn’t the prOAS (prepay-risk-and-option-adjusted spread) method require measuring prepay model risk and rewarding investors accordingly? Actually, the entire math of prOAS works when a prepay model is tuned from its objective (physical) to a risk-neutral stage. Since AD&Co prepay models always come with “dials,” they can be made risk-neutral using MBS benchmark prices. After this is done, one can run the

>>>