|
|
Post by GSV3MIaC on Jun 6, 2014 15:55:21 GMT
This pair have come up out of order .. Any guesses? They aren't really late ENOUGH to have been on the whole loans queue and bounced .. Maybe just finger trouble in the numbering department? I' be seen gaps before, but not, afaicr, retrograde numbering.
|
|
chrisf
Member of DD Central
Posts: 224 
Likes: 67
|
Post by chrisf on Jun 6, 2014 17:56:35 GMT
This pair have come up out of order .. Any guesses? They aren't really late ENOUGH to have been on the whole loans queue and bounced .. Maybe just finger trouble in the numbering department? I' be seen gaps before, but not, afaicr, retrograde numbering. Interesting. Notice how 6403 and 6405, just like 6394 and 6397 the day before, are 24 hours displaced from where they would be in numerical order. Could it be that 'whole loans' are only available for 24 hours before being unleashed on the likes of us?
|
|
|
|
Post by GSV3MIaC on Jun 6, 2014 19:10:40 GMT
Yep, I suspect that is the case, from replies over on the other forum. David said there were 2, maybe it is now 4. Eventually the goal is 3 hours and then open market. I'm still not too thrilled at the concept of getting second pick of what the HNWIs or companies didn't want.
|
|
chrisf
Member of DD Central
Posts: 224
Likes: 67
|
Post by chrisf on Jun 6, 2014 20:02:17 GMT
Just can't tear yourself away from 'the other place' then.
|
|
|
|
Post by GSV3MIaC on Jun 6, 2014 20:21:53 GMT
Seemed rude not to acknowledge the one FC reply I've had all week!
|
|
blender
Member of DD Central
Posts: 5,719
Likes: 4,272
|
Post by blender on Jun 6, 2014 20:26:34 GMT
It's worth a read.
Apparently only two loans offered as whole loans were rejected and were listed as partial yesterday. There are 28 drawn down plus the two rejected and none are A+ or C-. We are assured that the selection is random. How likely is it to select 30 new loans at random and select no A+ or C- loans? It is not that I disbelieve, just that it seems almost infinitely improbable.
|
|
chrisf
Member of DD Central
Posts: 224
Likes: 67
|
Post by chrisf on Jun 6, 2014 20:51:31 GMT
My simple(-ish) maths says 98.6% unlikely
|
|
chrisf
Member of DD Central
Posts: 224
Likes: 67
|
Post by chrisf on Jun 6, 2014 20:52:48 GMT
so not as bad as infinitely unlikely, just unlikely.
|
|
blender
Member of DD Central
Posts: 5,719
Likes: 4,272
|
Post by blender on Jun 6, 2014 23:06:27 GMT
It is not infinitely improbable, just very. To give all you need for calculation:
Distribution of the 30 whole loans. A+ 0, A 8, B 13, C 9, C- 0.
Distribution of the last 100 loans drawn down: A+ 10, A 24, B 28, C 26, C- 12.
From that the probability of not drawing a C- or A+ (specifically) at one try must be 0.78, and presumably the probability of not drawing a C- or A+ in 30 tries must be (possibly shows ignorance when as a young porker could have done it easily) 0.78 to the power 30, which I make less than 0.001. Any statisticians out there?
At that level it is perhaps easier to explain the result by an FC software coding error than by chance.
|
|
|
|
Post by bracknellboy on Jun 6, 2014 23:29:13 GMT
assuming you've done the maths right, then the odds would be less than you've calculated. Reason: they are drawing from a bounded pool not an infinite pool. Don't know how long the trial has been, but for example, if its 30 drawn from a potential pool of 100 then :
>>Distribution of the last 100 loans drawn down: A+ 10, A 24, B 28, C 26, C- 12.
[assuming for the moment that the 100 includes the 30] The chances of first loan you pick of the 30, (any one, but the first one you pick), not being A+/C- is 1- 22/100 which as you rightly point out is .78. However, its wrong to then extraploate to .78^30
The chances of the second loan not be not being A+/C- is 1-22/99: becuase its a finite pool and the pool has been shrunk by one loan which de facto is not an A+/C-, so only the divisor decrements
So you have a series which is (1-22/100)*(1-22/99)*(1-22/98) etc.
Continue the series for the total number of whole loans. Adjust the numbers according the atcual number of whole loans and actual drawdown and distrbution for the period.
By the way, this can be converted into something more easy to calculate by a bit of algebra, but since I left that stuff behind 30 years ago it would require a bit more thought, a bit less wine, and a bit of googling to get you that. (or at lesat to rummage in my bookshelves and find one of me old stats books: but working it out first principle would be more fun).
Of course if the actuality is that the set of whole loans is very small compared to the total set of loans (probably a couple of orders of magnitude) then the results will be ~the same. but if 30 and 100 are even remotely close to the actual ratios then I would expect quite a reasonable difference in the result
|
|
blender
Member of DD Central
Posts: 5,719
Likes: 4,272
|
Post by blender on Jun 7, 2014 6:46:09 GMT
Thanks Bracknell Boy. You are right that it would be different if the pool was limited. My fault for not explaining the 100. I only used that to find the probability of the next loan to be listed not being an A+ or C-. There were many more than 100 loans in May, but that is not the pool. Each new loan that comes up may be designated either whole or partial and each new loan request, whether whole or partial, is fresh and has that same .78 probability of not being an A+ or C-. What has gone before is irrelevent, like tossing a coin. Not getting an A+ or C- in a random selection of 30 new loans strongly suggests that something is not quite working right in the selection process.
|
|
|
|
Post by transo on Jun 7, 2014 14:58:43 GMT
So do these rejected whole loans make clear that they were previously listed as whole loans before being offered as partials. Looking back at my original query to FC on this I didn't ask that question, but noted that the FAQ said this would be made clear, and queried how it could possibly be made clear to autobidders. If there's no explicit indiction (and odd numbering sequence certainly doesn't count) then FC appear to not be doing what they said they'd do.
|
|
blender
Member of DD Central
Posts: 5,719
Likes: 4,272
|
Post by blender on Jun 7, 2014 16:14:48 GMT
These two have been identified by a very reliable and knowledgeable contributor IMO on the other inferior forum, as 6394 and 6397. These identities have not been confirmed by FC, only that there are two, nor can I find any reference to the whole loans on the loan requests. But are you sure that FC said they would identify the partial that had been whole at the trial stage? I would be surprised since the idea was for the trial to be private, except through the loan book.
GSVetc has asked the question about the randomness of trial risk bands on the other forum. I doubt there will be a speedy reply, but we will see.
|
|
|
|
Post by bracknellboy on Jun 8, 2014 8:09:12 GMT
Thanks Bracknell Boy. You are right that it would be different if the pool was limited. My fault for not explaining the 100. I only used that to find the probability of the next loan to be listed not being an A+ or C-. There were many more than 100 loans in May, but that is not the pool. Each new loan that comes up may be designated either whole or partial and each new loan request, whether whole or partial, is fresh and has that same .78 probability of not being an A+ or C-. What has gone before is irrelevent, like tossing a coin. Not getting an A+ or C- in a random selection of 30 new loans strongly suggests that something is not quite working right in the selection process. Blender. Have no doubt we are on the same page. Yes I had assumed you were referring to the actual set of loans over the period (albeit rounded numbers), not just the prior distribution or distribution within that book. Its always about finding the simplest but still appropriate (i.e. good enough approximation to the real world scenario) model. The straight forward p^n is perfectly good (in fact correct) way of 'future predicting' the probability of a particular outcome. However if you are looking at an outcome that has already happened, you have more information to go on, obviously i.e. the actual distribution within the set of loans from wihch they were sampled. Your comment on what has gone before being irrelevent is absolutely correct if you are simply working on the basis of historic distribution to get probability of particular future outcome: because one is allowing for a randomness in the banding of the stream of loans e.g. allowing for there being anything from 0 to 100% of each of A+/A...... in the set of loans from which the whole loans are chosen, as well as the randomness of selection from them. Once you have a defined data set (e.g. an actual set of loans from which the sample was taken) the 'each event is independent of what went before' is no longer relevant, because you are calculating the probability of the actual outcome from a known data set, in which case the combinations approach is the more accurate. However the p^n is still the preferable (i.e. easiest or even feasible to calculate) provided it remains a good enough approximation. As an example, going back to example of 30 from 100. If the actual set of 100 loans from which the sample of 30 was chosen contained 25 A+s, the predictive approach p^n would still give a finite (albeit small) probability of the 30 containing only A+s, which is patently absurd, whereas the combinations approach will yeild a probability of 0 for any outcome containing 26 or more A+s. This of course is because the first approach is still allowing for (probabilistically weighted) all potential outcomes for the distribution of bands within the actual 100. Looking at the actual loan book, and assuming I've extracted the data correctly, what I see is 28 whole loans and a total of 263 loans during the period (first whole loan onwards). The actual distribution of bands within that is: A (Low risk) 34.60% A+ (Very low risk) 11.41% B (Below average risk) 25.48% C- 9.13% C (Average risk) 19.39% so 20.53% A+/C- (54 from 263) Given that we are talking 28 from 263 then 79.5% ^ 28 is absolutely going to be the right calculation to do (super set large enough, sample size small enough relative to). Looking at the data further is interesting. The distribution within the whole loans is: A (Low risk) 25.00% B (Below average risk) 42.86% C (Average risk) 32.14% The distribution in the set of all A/B/C only is: A (Low risk) 43.54% B (Below average risk) 32.06% C (Average risk) 24.40% But of course at this point (going down to band level) the number of whole loans by band is becoming quite small: A (Low risk) 7 B (Below average risk) 12 C (Average risk) 9 So I think we can conclude that it is exceedingly unlikely that the whole loans are being randomly selected from all loans, and highly probable that they are only being selected from loans excluding A+/C-; OR that they are being selected but that the criteria used by the whole loan buyers is subsequently rejecting all A+/C- (i.e. either way a hard no completion of A+/C-). Its also likely that the actual distribution of whole loans is compatible with a random selection from just A/B/C loans, but without effort difficult to know whether its a reasonable outcome (given the sample size): it could also be (likely?) that the underweight in A may have more to do with acceptance/rejection of whole loans by the buyers rather than any non-randomness in original selection.
|
|
blender
Member of DD Central
Posts: 5,719
Likes: 4,272
|
Post by blender on Jun 8, 2014 9:07:40 GMT
Thanks BracknellBoy. Yes we are exactly on the same page. I took the 30 loans rather than the 28, but it is true that FC has identified only the 28 as ABC so your calculation is safest. I think we agree that the chance of drawing 28 loans for the trial, with no rejects, as A,B or C is less than one in a thousand. So it is right to test the statements that the selection was random over all loans and/or that no loan requests were rejected from the random selection. It is also the fact that it was the outlying risk bands which did not appear, one having low interest after FC's take, the other having the highest risk, which causes questions.
Personally it would not trouble me too much if the trial had been constrained, with our knowledge, to suit the needs of the triallist(s), as long as the substantive offering that followed was random. But I do like to be confident that I can understand correctly what my chosen platform is saying to me.
|
|
