Lyft Rides Analysis | Kevin Chen

ANALYZING IMROVEMENT OPPORTUNITIES FOR LYFT RIDES

Analyzing historical data of 150,000+ Lyft rides from public sources to gain insights into product performance and behavior.

Dataset: Lyft Data Challenge (Kaggle)

Technologies and Methods Used: Python (pandas, numpy, matplotlib)

INTRODUCTION

In the ever-evolving landscape of ride-sharing services, Lyft has established itself as a prominent player. With However, Lyft needs to continue growing its revenue, and needs to have both drivers and users on the platform to meet supply and demand needs. To delve deeper into the Lyft experience for both riders and drivers, as well as identify potential avenues for growth, a comprehensive analysis of the available data becomes crucial. By examining key metrics such as successful pick-up rates, prime-time dynamics, driver behavior, and peak ride times, this article explores valuable insights that shed light on rider satisfaction, driver retention, and strategic opportunities for Lyft in the competitive ride-sharing market.

DATA PREPROCESSING

I obtained the data from Kaggle and is known as the Lyft Data Challenge. Before proceeding with analysis, I first checked to see if there were any extraordinary observations in my data. My data consisted of three separate files, which I joined together using inner joins on a common key:

driver_ids : contains the driver id's and onboarding dates

ride_ids : contains information about rides including ride distance and duration, ride id, the prime time charge, and the associated driver id

ride_timestamps : contains the time of event, ride id and event(requested, accepted, etc.)

On looking at the data, there was only one NA value in the timestamp file
no odd dates, times, or events
The starting date of this data was March 27 2019 and the end date was June 27 2019.
There were ride distances of less than 0 and as high as 700,000 units. Knowing Lyft rides cannot exceed 100 miles, I assumed that this distance was in feet. Thus, I eliminated entries with more than 100 miles or 528000 feet

WHAT'S IMPORTANT, BASED ON THE DATA

I thought the following insights could be important given the data:

% of successful pick-ups after acceptance
Mean difference between request and accept times per group of prime time
Drivers with least number of completed rides
Are number of rides increasing/decreasing/stagnant over this duration?
Mean number of rides per driver on given day of the week
How many drivers with at least one completed ride in April also completed a ride in June?
Frequency of Lyft rides according to hour and day

ANALYSIS

PERCENTAGE OF SUCCESSFUL PICK UPS

Firstly, I wanted to measure the percentage of successful pickups. I don't really know the standard of "good" and "bad" in the industry, and so was curious to find out. However, I found that the successful pick-up rate was 100%; from just my personal experience, I know that some riders do not get picked up for a variety of reasons, and this led me to believe that this dataset consisted only of successful, complete rides.

I then wanted to find the mean difference between accepted and requested time across prime-time groups. As mentioned earlier, higher prime-time rate could be associated with higher average time between requests and acceptances, but the reverse could also happen depending on the situation. The truth seemed to lie somewhere in between the two situations.

A graph of mean difference in times across prime time groups revealed an almost normal distribution, with mean time difference increasing by prime time group until 150 before falling. Interestingly, there were two outliers in the data, with mean time difference for 300 and 400 prime-time groups showing a very high interval. However, on looking at the data, this anomaly was caused primarily by one outlier in each of these groups. For the 350 Prime-Time group, one of observation had a wait time of 1216 seconds, while the 450 Prime-Time group was skewed by a ride with >422 seconds between request and acceptance

DRIVERS WITH LEAST NUMBER OF COMPLETED RIDES

I made a table of drivers with least number of completed rides in the timespan of this dataset. The mean number of rides completed by drivers during this time was 219. On the other hand, the top 10 drivers in terms of least rides completed had less than 25. Identifying these drivers could be useful for Lyft to zero in on why they might have driven so infrequently and whether or not they left Lyft entirely. This could in turn be used to collect survey data to improve driver experience or initiatives to lure them back to Lyft.

DRIVER BEHAVIOR

Continuing on the topic of Lyft drivers, I did further analysis to uncover more about Lyft driver behavior during this time frame.

First, I wanted to see if the number of rides were increasing or decreasing with time. I eliminated March since observations start only on March 27. The results did show a favorable trend for Lyft, with a huge spike in rides from April to May. There is a slight decrease from May to June, but this could be attributed to the the dataset ending its time period on June 27. Correlation is not causation, however, and a number of factors could have contributed to this spike (seasonal, events, transport regulations), and the 10,000 ride deficit from May to June is also unlikely to be coverd in the three remaining days in June

A healthier measure of drivers' affinity and happiness with Lyft would be retention. Monthly riders is Lyft's North Star metric, and on the driver side, I presume Lyft considers this metric crucial for gauging driver satisfaction. The retention rate from April to June for drivers was 78.06%, meaning 78% of drivers who completed at least one ride in April completed one ride in June. This amounts to a churn rate of about 22%, which does seem quite high to me.

Indeed, the number of unique drivers spikes to about 800 in May, but dips to below 800 for the month of June. However, as mentioned, although the number of returning drivers decreased, the number of rides increased. One positive is that those current drivers have a larger share of the rides, but fewer drivers could result in deficit of drivers during peak hours.

PEAK TIMES

Peak ride times could be measured in number of rides per day. To gain a better sense of ride-hailing trends, I calculated the frequency of Lyft rides by day and per hour of day. This analysis returned the resulting graph:

The resulting graph was almost as expected, with peak ride-hailing frequencies for Lyft occurring late at night on weekends and the beginning and end of working hours on weekdays. However, the heat map does suggest that there are slightly more Lyft rides occurring at the end of work than at the beginning of the day. This information could be useful for Lyft for crafting plans to meet demand during peak hours and also possibly to raise demand and supply of rides during other times such as the beginning of a working day when people are going to work (a daily carpool feature, perhaps?).

WHY COULD THESE INSIGHTS BE IMPORTANT?

Percentage of successful pick ups could be an important indicator for rider retention. Low or drop in percentage of successful pick ups after accepting a request could turn riders away from Lyft and towards competitors
Prime time is when Lyft prices are higher than normal. One might expect the mean request and accept times during prime-time to be lower than during non-prime-time since drivers might be more eager to accept a costlier ride. On the other hand, this metric could also be misleading since there is more demand during prime-time, leading to increased pressure on the driver supply which could be outweighed by demand, leading to longer average accept times. However, in general this would probably be something Lyft would be looking to decrease all the time (faster acceptance, greater rider satisfaction)
Since this dataset gives us a lot of information about drivers, driver data could be used to see whether drivers are enjoying being Lyft drivers compared to competitors. Ride-hailing services are a two-way street: if demand outweighs supply, than riders might turn elsewhere for transportation. If supply is an issue, then Lyft might be interested in isolating this as a problem.

OPPORTUNITIES FOR FURTHER RESEARCH

While I enjoyed writing this article and conducting analysis, there are still ways to capitalize on this data and do even more with more information. For example: - Mean distance per prime-time group. This could help pricing strategies for Lyft - Predicting destination: - If location data were available, data mining methods could help predict which destinations are most common according to pick-up location and time of day/day of week. This could help in devising strategies for pricing and also to make sure demand for drivers is met accordingly.