Rennpoints predictions
04-20-2015, 09:22 PM
#16
Rennlist Member
The prediction algorithm is a bit involved...
It started with something that a buddy and I noticed a long time ago that our lap times relative to the class records tended to be very consistent. I.e, if my lap time at a track is 1.05x the class record, it is likely to be 1.05 times the class record at most other tracks (usually within 2/10s of a percent).
In the first incarnation of coming up with predictions, I tried to apply that formula and ran into a few basic problems.
1) If an event is lightly attended and a "slow" driver puts up a class record due to poor competition showing up, the prediction for that driver tends to greatly overstate how good they will do at future events.
2) Since the comparison is made against the track record, the best that you could be predicted to do is as good as the track record.
3) If your first couple events are at rainy events, it will generally give you poor predictions as it sees that you are driving slow (compared to the class records)
I thought about this for a while and came up with a new comparison mechanism that resolved a few of those issues
1) The algorithm looks at all sessions that you ran against other people in your class and that have also raced at the track that we are trying to predict. It assumes that the ratio between your lap times will stay constant. For example, if at another track, you turn a 61 second lap, and your peer turned a 60 second lap, it assumes that ratio stays constant. So, if your peer turned a best of a 120 second lap at the prediction track, it assumes you will turn a best of 122 seconds.
By comparing racers at the same session, it nicely solves the problems of weather conditions (as both people see the same conditions). It also allows the prediction to be faster than the standing class record.
Now, the problem remaining is which driver to compare against and which of your sessions should it use as the base? I.e., we want to generate a prediction based on when you were running a good lap time (e.g., no car issues) and your peer was also running a good lap time. For that, we compare your lap time against the class record and your peer's time against the track record. Also, to ensure that your competitor is consistent, we compare his best lap time at the prediction track against the track record. We look for instances where all three comparisons are close to the track record.
As a formula, we are trying to maximize the following:
(class record at comparison track)/(your lap time at comparison track) * (class record at comparison track)/(your peer's lap time in same run session at comparison track) * (class record at prediction track / your peer's best lap time at prediction track)
Once we find that number, we then look for other peers within 2% of that product. We then use the ratio of lap times to come up with predictions and average them all together.
I've found that the prediction is remarkably accurate but am aware there are a few flaws:
1) If somebody did really well a long time ago, it still keeps coming up with good predictions for that person. I should have something that ages out old results
2) Class records are always based on the current class record. It may be worthwhile to use the class record as it stood at the time (that might also solve the above problem)
Hope that helps clear things up...
It started with something that a buddy and I noticed a long time ago that our lap times relative to the class records tended to be very consistent. I.e, if my lap time at a track is 1.05x the class record, it is likely to be 1.05 times the class record at most other tracks (usually within 2/10s of a percent).
In the first incarnation of coming up with predictions, I tried to apply that formula and ran into a few basic problems.
1) If an event is lightly attended and a "slow" driver puts up a class record due to poor competition showing up, the prediction for that driver tends to greatly overstate how good they will do at future events.
2) Since the comparison is made against the track record, the best that you could be predicted to do is as good as the track record.
3) If your first couple events are at rainy events, it will generally give you poor predictions as it sees that you are driving slow (compared to the class records)
I thought about this for a while and came up with a new comparison mechanism that resolved a few of those issues
1) The algorithm looks at all sessions that you ran against other people in your class and that have also raced at the track that we are trying to predict. It assumes that the ratio between your lap times will stay constant. For example, if at another track, you turn a 61 second lap, and your peer turned a 60 second lap, it assumes that ratio stays constant. So, if your peer turned a best of a 120 second lap at the prediction track, it assumes you will turn a best of 122 seconds.
By comparing racers at the same session, it nicely solves the problems of weather conditions (as both people see the same conditions). It also allows the prediction to be faster than the standing class record.
Now, the problem remaining is which driver to compare against and which of your sessions should it use as the base? I.e., we want to generate a prediction based on when you were running a good lap time (e.g., no car issues) and your peer was also running a good lap time. For that, we compare your lap time against the class record and your peer's time against the track record. Also, to ensure that your competitor is consistent, we compare his best lap time at the prediction track against the track record. We look for instances where all three comparisons are close to the track record.
As a formula, we are trying to maximize the following:
(class record at comparison track)/(your lap time at comparison track) * (class record at comparison track)/(your peer's lap time in same run session at comparison track) * (class record at prediction track / your peer's best lap time at prediction track)
Once we find that number, we then look for other peers within 2% of that product. We then use the ratio of lap times to come up with predictions and average them all together.
I've found that the prediction is remarkably accurate but am aware there are a few flaws:
1) If somebody did really well a long time ago, it still keeps coming up with good predictions for that person. I should have something that ages out old results
2) Class records are always based on the current class record. It may be worthwhile to use the class record as it stood at the time (that might also solve the above problem)
Hope that helps clear things up...
