This project analyzes hourly bike-share usage data to identify key demand drivers and guide decisions across product, operations, and marketing teams.
Stakeholders:
- Product Manager: Rider behavior & feature testing.
- Operations Lead: Staffing, rebalancing, maintenance planning.
- Marketing Lead: Promo timing, rider segmentation.
- Policy & Ethics Advisor: Ensure equitable access and responsible insights.
- Cleaned and transformed data using pandas in Python.
- Visualized hourly ridership trends by time, weather, and user type.
- Peak hours are 8 a.m. and 5 p.m. with roughly 350 & 475 average hourly riders repectively
- Rindership increases before and decrease after theses points
- Ridership has a strong relation ship with people commuting to work
- Less clear weather correlates with lower average ridership
- Clear weather has >200 Average riders compared to heavy snow/rain with < 100
- Most riders are from registered users with 81.2% comapared to casual users making up the remaining 18.8%
- People who rode a bikes are likely to be repeat users so they register
- Hypotheses:
Null hypothesis (H0): Mean hourly rides on working and non-working days are equal.
Alternative hypothesis (H1): Mean hourly rides on working and non-working days differ.
- Test: Welch’s t-test
- Result: p-value = 0.00004
- 95% CI: (6.15, 17.45)
- Decision: Reject null hypothesis!
- Practical Insight: There is a statistically and practically significant difference in hourly ride counts between working and non-working days. On average, working days see 6 to 17 more rides per hour, which can translate to hundreds of additional daily rides. This pattern reflects likely commuter behavior and should guide weekday-focused operational planning and targeted pricing or promotional strategies.
Hypotheses:
Null hypothesis (H0): All seasonal means are equal (no difference in average hourly rides).
Alternative hypothesis (H1): At least one seasonal mean is different.
-
Test: One-way ANOVA
-
Post-hoc: TukeyHSD
-
Decision: Reject the null hypothesis!
-
Insight: Average hourly rides do significantly differ between at least some seasons. The size of these differences with such a small p-value have clear operational, financial, and strategic implications for pricing, resource allocation, and customer engagement. making it not only statistically significant but practically significant aswell.
-
The appropriate post-hoc approch would be a Tukey's HDS for pairwise comparisons between seasons to determine which specific pairs of seasons differ significantly. This would allow stakeholders to focus on seasons that may be underperforming.
Objective:
Measure the impact of an app feature (launched on 2012-09-01) on weekday evening (17:00–19:00) ridership during good weather.
Eligibility Criteria:
workingday == 1hr ∈ {17, 18, 19}weathersit ∈ {1, 2}hum ≤ 0.70
Windows:
- Pre (Group A): 2012-08-04 → 2012-08-31
- Post (Group B): 2012-09-01 → 2012-09-28
Design:
Data was stratified by (weekday × hour) and matched by truncating to equal sample sizes across time slots. This ensured balanced groups and fair comparison.
Balance Check
Test:
- Paired t-test
- Null: No difference in avg. hourly rides
- α = 0.05
Results:
- p-value: 0.0006
- 95% CI: [-63.6, -18.7]
- Mean Increase: +41.15 rides/hour
- Cohen’s d: 0.30 (small–medium)
✅ Statistically and practically significant (~15% lift).
Suggests increased demand likely due to the feature — may warrant operational changes (e.g., staffing, fleet availability).
- Invest in feature development that supports commuter behavior
- A/B test showed a +41.15 rides/hour increase after feature launch.
- Prioritize features aimed at weekday evening rides (17:00–19:00)
- Support the registered user base
- With ~81% of rides from registered users, optimize for retention (e.g., loyalty rewards, referrals)
- Run additional experiments
- Continue controlled A/B testing to evaluate product changes over time
- Prioritize commuter peak hours
- Focus staffing, bike rebalancing, and maintenance during 8 a.m. and 5 p.m. peaks.
- Implement weather-based resource planning
- Clear weather = high demand; adjust fleet and station support accordingly
- Plan for seasonal fluctuations
- Use seasonal trends to shift operational focus (e.g., more bikes in summer/fall)
- Target weekday commuters
- Use promotions and messaging during high-traffic windows (morning/evening)
- Design seasonal campaigns
- Create special promotions in underperforming seasons identified via ANOVA results
- Leverage weather
- Run weather-sensitive promotions
- Doesn't include pricing model
- Dates only range from 2011 - 2012
- season date range is incorrect
- dont over prioitize peak hour commuters
- dont over prioritize registered users
slides/: presention with findings and insights.data/: Cleaned and raw datasets.notebooks/: Jupyter notebooks for EDA, testing, and modeling.README.md: Project overview and conclusions.