I have had a weird theory for years that Uber fits their ETA model using mean absolute error regression instead of ordinary least squares. Since wait-times have a long right tail, this will cause a bias that underestimates mean wait times (median < mean).

@bernhardsson We used quantile regression at Instacart to have control over where in the distribution we predicted, and then A/B tested our way to the right q

There are two types of ETAs: - pre-booking ETA which is an acceptance-likelihood weighted of nearby drivers. Bias here is driven by the bias of the driver acceptance model. Very hard to get right and constantly get calibrated. - post-booking ETA which is classic time between two locations on the map

@bernhardsson the explanation was that the wait times are calculated by a team that has nothing to do with dispatch and routing in most markets.

@bernhardsson I have been telling my friends this forever and no one believes me, especially on the new rideshare app Empower. Seems like straight line distance calculation half the time

@samxpatterson Arrival times have a long tail which means the median is lower than the mean

I have had a weird theory for years that Uber fits their ETA model using L1 regression instead of ordinary least squares. Since wait-times have a long right tail, L1 reg will cause a bias that underestimates mean wait times.

@bernhardsson Lasso shrinks the coeff not the response tho?

@bernhardsson I'm 100% sure they've got internal data proving that more accurate, less optimistic ETAs are revenue negative

@bernhardsson Nah I bet they have a perfectly accurate model of the probability distribution of wait times, but they've surely A/B tested that showing the p25 wait time instead of the mean or median is +pnl

@bernhardsson it’s a conscious decision in that the feedback mechanism is pretty weak; what are you gonna do, not trust uber etas the next time?

@bernhardsson Or more accurately that model was selected as a legally defensible way to create reliably biased wait times

@can It’s one thing if the prediction is bad (as in, unbiased but high variance). But the reality seems to be that it has a big consistent bias. Which feels kind of sinister imo. Though my theory would rationalize it a bit

@bernhardsson why would lasso cause a bias towards shorter wait times? I don't think it's inherent that lasso would be biased towards that but maybe there's something else I'm not thinking about?

@bernhardsson Very few people actually take Ubers in multiple markets but they are vocal. I think it’s more that ETA is generally a pretty weak factor in decision. You first decide if you are gonna do ride share and decide between lyft/uber based on price.

@bernhardsson AFAIK, Lasso is a biased estimator (because of the L1 regularizer) but you cannot predict the consistent sign of the bias by definition, no? Unless I misunderstood your argument.

@can Ok but why does the eta estimation seem biased towards shorter wait times in most cities?

@can Though it’s possible the bias is consistent per market in a way that it cancels out globally?

@F_Vaggi I thought L1 regularization is applied to coefficients, L1 regression is applied to residuals?

For the first ~decade or so after founding, Uber had quite accurate ETAs. At some point around 2 years ago they became consistently biased towards underestimating the wait time. I imagine that is when a team internally realized there is no penalty for lying and no incentive to not lie.