Yank model: Challenger to the Duckworth/Lewis model

[yashsr]

Hello everyone,

I and a friend of mine have developed a mathematical cricket model known as 'Yank Model' and have challenged the Duckworth/Lewis method. You can check our website at:
Yank's Model

We've covered plenty of examples in T20 and ODIs in our Comparative Analysis. We've also showed why our model is superior to the D/L and the VJD method and have also pointed the apparent flaws and inconsistencies of the 2 models. We also have done a real match simulation on the Champions League 2010 matches. But we still haven't put up the details on the computation part as we've not completed the process of patenting.

I'd like the cricket fans from CricketWeb Website to have a look at our model and give us a feedback. I'd urge the readers to just see the situation and guess your own reasonable target before checking the targets of the 3 models - Yank, D/L and VJD, and then judge which is better. I'd urge readers to think in a cricketing manner and not just mathematically. As always, constructive criticisms and comments are welcome! Although we've put up our own comments regarding almost every example, you can always have a different opinion which we'd love to know. One thing I'd like to tell everyone is that I personally believe that our model has no or very little flaws because we've worked for about 4 months and identified flaws in the D/L and the VJD model and created a model accordingly based on real match data. But you're more than welcome to challenge this claim!

Thanks,
Yash

 

[NUFAN]

Brilliant work! It obviously looks like you've done your research and the examples were very easy to follow and the scores for the Yank model look like they are more realistic.

 

[Athlai]

Yeah looks pretty decent.

 

[Top_Cat]

Is any of your research published in a peer-reviewed journal? Would like to have a read.

 

[yashsr]

Is any of your research published in a peer-reviewed journal? Would like to have a read.

We've just approached a few journals and it takes a long time for them to publish it....so guess you'll have to wait!

@everyone: Thanks for the feedback!

 

[Top_Cat]

Well aware, trying to get a paper up in Addiction myself.....

 

[yashsr]

Well aware, trying to get a paper up in Addiction myself.....

LOL....Any ideas if we can give our model to more than 1 research journal at a time??? Someone told us that we cannot give it to more than 1 reserach journal...is it true?

 

[honestbharani]

link taken off or what?

 

[yashsr]

link taken off or what?

Yank's Model is still working

 

[Athlai]

Would like to see your model applied to more matches where D/L came into play. All of them if it at all possible. Would make interesting reading.

 

[Uppercut]

Fantastic effort this, congratulations. Completely agree with some of your criticisms of D/L. In particular the case study where New Zealand could be 29/1 after 8 overs chasing 85 in a T20 game and be deemed the losers was ridiculous.

 

[Quaggas]

LOL....Any ideas if we can give our model to more than 1 research journal at a time??? Someone told us that we cannot give it to more than 1 reserach journal...is it true?

Submitting to two journals at the same time is a big no-no.

 

[Top_Cat]

Yup. Big no-no. Fact of research life.

 

[yashsr]

Submitting to two journals at the same time is a big no-no.

Ok thanks....so I guess we'll have to wait

 

[yashsr]

Would like to see your model applied to more matches where D/L came into play. All of them if it at all possible. Would make interesting reading.

Yeah we're planning to add more examples in the near future along with some real match simulation on other major tournaments like World Cup etc. also...

@uppercut: Yes true.... D/L particularly fails in low scoring matches as per our research....

 

[Goughy]

If you are wanting honest feedback-- its great to try to continuously improve systems and processes we have in place but I do have a few points about his page which I only offer as you requested feedback.

CLT20 Simulation Yank's Model

Firstly it seems a little strange using D/L for the method you are. Im not convinced it is designed to be used to predict future totals.

Secondly, the data you present has the D/L method closer to the actual score more often than yours 12 vs 10. Not sure how that helps your cause.

Thirdly, your conclusion that your system is more accurate based on the fact that across all the games your system more acurately predicted the total aggregate number of runs than D/L is meaningless.

Consider this example
Game 1: Actual runs- 100 Predicted runs- 200
Game 2: Actual runs- 200 Predicted runs- 100
Game 3: Actual runs- 150 Predicted runs- 100
Game 4: Actual runs- 100 Predicted runs- 150

Total actual runs = 550
Total predicted runs = 550

Perfect system then? No, obviously the totals mask the fact that for each game the prediction can be way off. That makes the method you used to draw a conclusion dubious.

Fourthly comments like " CSK made a good recovery from a bad start", "Guyana were expected to score more", "Victoria were expected to score more, because they had wickets in hand and probably set batsmen batting", "MI made a miraculous recovery ", "Central Districts made a good recovery considering the poor start.", "Wayamba’s batting order just collapsed. No projection method can account for such a collapse.", "South Africa made a astounding recovery from a relatively weak position ", "CSK just failed to accelerate in the last 10 overs.", and "Warriors failed to take advantage of a relatively solid start" would seem to suggest that a lot of variables are ignored and the system predicts based on what is expected to happen based on previous events rather than what can happen in a game full of variables.

 

[yashsr]

If you are wanting honest feedback-- its great to try to continuously improve systems and processes we have in place but I do have a few points about his page which I only offer as you requested feedback.

CLT20 Simulation Yank's Model

Firstly it seems a little strange using D/L for the method you are. Im not convinced it is designed to be used to predict future totals.

Secondly, the data you present has the D/L method closer to the actual score more often than yours 12 vs 10. Not sure how that helps your cause.

Thirdly, your conclusion that your system is more accurate based on the fact that across all the games your system more acurately predicted the total aggregate number of runs than D/L is meaningless.

Consider this example
Game 1: Actual runs- 100 Predicted runs- 200
Game 2: Actual runs- 200 Predicted runs- 100
Game 3: Actual runs- 150 Predicted runs- 100
Game 4: Actual runs- 100 Predicted runs- 150

Total actual runs = 550
Total predicted runs = 550

Perfect system then? No, obviously the totals mask the fact that for each game the prediction can be way off. That makes the method you used to draw a conclusion dubious.

Fourthly comments like " CSK made a good recovery from a bad start", "Guyana were expected to score more", "Victoria were expected to score more, because they had wickets in hand and probably set batsmen batting", "MI made a miraculous recovery ", "Central Districts made a good recovery considering the poor start.", "Wayamba’s batting order just collapsed. No projection method can account for such a collapse.", "South Africa made a astounding recovery from a relatively weak position ", "CSK just failed to accelerate in the last 10 overs.", and "Warriors failed to take advantage of a relatively solid start" would seem to suggest that a lot of variables are ignored and the system predicts based on what is expected to happen based on previous events rather than what can happen in a game full of variables.

Thanks for your feedback, Yes you're correct, we haven't put up absolute totals. The thing we're trying to prove with this simulation is that D/L almost always gives low projected scores which means its acceleration level for T20 is very very low as compared to what happens in an actual game.

It is really impossible for a model to come very close to real match scores because of matches like:
Match 17) South Australia scored 191 from a relatively bad position of 68/3 in 10 overs

Match 16) Wayamba score 106 from a very healthy start of 81/2 in 10 overs

When we take normal differences(not absolute), we can find out whether the acceleration level of the model is simulating what happens in real matches. We initially planned to take absolute differences but it wouldn't prove any point because of bizarre matches.

Readers are actually requested to look at the 10 over score of the team and project the 20 over score for their team according to their own judgement and then have a look at Yank and D/L Model Projected Scores rather than only comparing with the actual score attained in the match. Using the actual score of every match wouldn't be the best way to judge, using a total of actual score would a far better measure because of extreme matches....

As for considering other variables, we cannot take into account variables like quality of the opposition, skill of the players etc.... If we were to do that, teams like Zimbabwe and Bangladesh would almost always lose a match because even if they're doing well at the 25 over stage, its very easy for them to fall out against the bigger teams and lose the match. A Model can only look at the score and predict the future(if interruption is in 1st innings) and then calculate par score and the target...

We’ve calculated Projected scores for all matches in the above tournament from the half-way stage for the Team batting first. Please note that although a model is not required to do this in the real matches, this calculation does happen on the back-end when an interruption in the 1st innings takes place.

Consider a case where in a match between Mumbai Indians(MI) and Lions in which Lions is batting first and is at 78/2 at the 10 over stage when it rains and only 10 overs can be bowled, so we need to calculate target for MI in 10 overs. For this, we need to calculate the 20 over par score(i.e. projected score) for Lions in its entire quota of 20 overs.

In the above case, Yank’s projected score is 168 and D/L’s Projected Score is 150. Accordingly, Yank target would be 102 and D/L target would be 88.

 

[yashsr]

I think we've focused a little to much on the Real Match Simulation...we'd love to have feedback on the general performance of the Yank Model as against D/L model and VJD model both in ODIs and especially also in T20..

 

[yashsr]

About yesterday's match which was the 2nd ODI in the series between SL and Aus:

SL 161/2 in 34 overs.
Match reduced to 45 overs.
SL 213/3 in 41.1 overs.
Innings ends and Aus get only 39 overs to bat.

39 over Target

Yank Method :240

D/L Method :244

VJD Method :231

Aus 139/5 in 27.4 overs
Match reduced to 38 overs.
Find target for Australia in 38 overs.

38 over Target

Yank Method :236

D/L Method :240

VJD Method :227

 

[Prince EWS]

This is a gun thread. It's disappointing that those who constantly complain about D/L without really understanding at all how it works or why it's the way it is haven't bothered posting in it. I've always been a fan but I think this does seem to produce slightly more realistic scores, particularly in T20.