Daniel Kagan
Data Science Fellows June 2020 Cohort
Abstract
Autofleet improves fleet performance by carrying out rebalancing moves which transfer vehicles from regions of oversupply to regions of excess demand. The algorithm currently used in production assumes the supply at each location is static and equal to the current supply while rebalancing tasks will be performed throughout the day. This assumption is obviously incorrect as clients move vehicles from one place to another when they use the fleet’s vehicles. I significantly improved the quality of the supply prediction relative to the baseline. This has the potential to greatly improve the company’s effectiveness by eliminating rebalances that are not needed or are actively harmful to the fleet’s performance due to shifts in supply.
Challenges (at least two)
- It was difficult to figure out whether ride bookings were a good predictor of supply – in fact they were very bad and figuring this out took quite a while.
- Some rebalances are used to return vehicles that had been dumped outside the territory of the client – I had to compensate for this to properly calculate organic shifts in supply.
Achievements (according to KPIs)
- Improved supply prediction over baseline by 30.8%
- Adapted the company’s simulation framework to make it possible to find the effect of the new supply on the company’s overall bottom line. This is incomplete, because the simulation must be redesigned to handle variable supply and take into account when rebalances occur (they are currently assumed to happen at midnight).
- Showed that local supply relative to its average value was a key feature of the model, which may allow for an improvement in demand prediction as well.
Future project development
First, the simulation framework must be redesigned to work with the variable supply predicted by the model. Secondly, the model should be tested using prediction intervals other than 1 day (that interval is based on the algorithm for demand prediction and may not be ideal for supply prediction). Finally the model can likely be improved with the addition of new features and possibly by more complex techniques like neural networks.
