e-n-f / ridership-model Goto Github PK

For the record, these are my predictions for the daily ridership of stations on the Silicon Valley BART extension, if the 7 stations currently planned or under construction are built.

The model predicts ridership of existing stations with a geometric standard deviation of 1.5, meaning we can be 95% confident that the ridership will not be lower than 44% of this prediction or greater than 225% of it.

I consider it somewhat unlikely (1 standard deviation above my prediction) that the three stations of the Berryessa Extension will have 23,000 riders per day at opening as predicted. I predict 4698 when Warm Springs alone opens, and 18,474 for the three stations together. Maybe by 23,000 they mean both entries and exits, in which case I am optimistic.

The 55,000 projected daily riders of the remaning 4 stations is more unlikely, 2 standard deviations above my prediction. Again, maybe they really mean both entries and exits, in which case it's only a little high.

If Livermore and eBART riders behave like other BART riders, I also predict the following:

The prediction comes from matching LEHD commute flows to existing BART station-to-station ridership counts with probabilities from the BART station profile study.

The stages are:

• Given all known Bay Area commutes, calculate the distance from their origins and destinations to the nearest BART stations or hypothetical future stations.
• Calculate a probability that this trip would be taken on BART based on the product of the lognormal curves from the Station Profile Study: 1260 feet (times or divided by 2.8) from work and 5682 feet (times or divided by 3.35) from home, divided by the curves for the total number of employed people who live or work at different distances from the nearest BART station.
• Multiply each probability empirically by the 0.7th power of the distance between the two stations, since longer trips appear to be more likely to be taken by BART.
• Scale the probability to match the actual station-to-station ridership counts: 0.1 times the 0.64th power of the probability.
• Sum the estimated station-to-station ridership count for each station pair
• For each station, scale it again to match the station totals by calculating 1.83 times the 0.74th power of the station total.

There are a lot of regression-to-the-mean problems along the way, but it's not too bad.