search

mikebryant / ac-nh-turnip-prices

Price calculator/predictor for Turnip prices
https://turnipprophet.io
Apache License 2.0
1.44k stars 245 forks source link

Prediction Inconsistencies #209

Open CampConn opened 4 years ago

CampConn commented 4 years ago

I was entering data for this week for a friend, and got this error: https://turnipprophet.io?prices=92.64.71..........

I was told that his memory isn't that great, and that the Monday morning price may be wrong, so I removed it to get this: https://turnipprophet.io?prices=92..71..........

What's strange is that all of the Small Spike patterns and about half of the Fluctuating patterns seem to indicate that 64 would've been a valid option for Monday morning, with the Small Spikes saying "37 to 83" and the Fluctuating ones saying "56 to 74." Is this an inconsistency in the calculations? Putting 64 for the morning without the 71 for the afternoon (https://turnipprophet.io?prices=92.64...........) shows 71 as being impossible for the afternoon, so is it just an issue with trying to retroactively predict previous prices?

theRTC204 commented 4 years ago

Seems like this is another example of the fudge factor causing the predictions to be wrong. We recently added some margin of error to the calculations, because we were seeing many reports of people unable to find matched patterns - the side effect of this unfortunately is that some patterns which are actually not possible are appearing in the results. Where previously we at times were returning to few results, now we are returning to many results.

There's some additional context in #198 and we have some work in progress to address this, but it will take a couple days.

mikebryant commented 4 years ago

@CampConn Your second hypothesis is correct - we only do forward-prediction, so that 37-83 range in the table/graph is only based on the Sunday price, not the Monday PM.

I don't think this is really worth putting time into - I personally only care about predicting forwards. (And time-travelling backwards resets everything anyway, I believe..)

marianhlavac commented 4 years ago

Is this fixed by the #219 ?

theRTC204 commented 4 years ago

Seems I was wrong to associate this issue with fudge factor - it is actually what mikebryant said; the application doesn't do backwards prediction.

What we are seeing here is.. if the Monday AM value was in fact 64, then 71 would not have been possible for Monday PM. However, if the Monday PM value of 71 is correct, then 64 on Monday AM could not have been possible.

One way or another, it appears that one of those values was incorrect.