Problem Statement

Trying to predict how bond auction result ( in terms of yield ) is different from its forecast (the when-issued yield ).

More info:http://www.mortgagenewsdaily.com/mortgage_rates/blog/242898.aspx

Data deficiency:

Auction only happens monthly. Therefore there are only 850 auctions result for bond of all maturity during the last 20 year interval.

If dividing the auction result by maturity term, it would leave each category with less than 200 results.


Plot of auction yield and when-issued yield difference , separated by maturity

enter image description here

Plot of auction yield and when-issued yield difference , not separated by maturity

enter image description here


Should I stay away from fitting a different predictive model for each bond maturity due to data deficiency ?

Instead, Should I do one-hot encoding and use bond maturity as a feature and then fit a single model for all 850 result ?

  • $\begingroup$ You should try this question on cross-validated too. My advice before anything else and certainly before torturing the data (the probleme is it will talk, eventually) is to start by splitting the data so that you keep a real uncontaminated holdout set. Then try fitting one or several models on the training set. You can probably set things up so that the number of different models to train is a hyperparameter and optimise for that. It may be that the optimal is neither one for all nor one for each, but perhaps one for “short” maturities and another for “long maturities” for example. $\endgroup$
    – Ivan
    Apr 1 '18 at 7:43
  • $\begingroup$ @Cloud: your chart implies WI-auction yield spread of up to 400,600bps. I think worth to check the data set and exclude anything over the steepness of 2y30y curve yield $\endgroup$
    – rrg
    May 25 '18 at 16:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.