Forecasting Demand by Weather: An Operations Audit

Summary

Managers often rely on intuition to adjust staffing during bad weather ("It's raining, cut the shift by 50%"). This project uses Hypothesis Testing to quantify the exact impact of rain on demand, providing a data-driven recommendation for dynamic rostering.

The Insight: Rain reduces demand by 38.7% (not 50%). Current manual adjustments are likely resulting in under-staffing and lost revenue.

---

The Business Problem

Context: A Micro-mobility operator in London. The Pain Point: Over-staffing on rainy days burns cash; under-staffing loses customers. The Goal: Determine the statistically optimal staff reduction percentage for rainy days.

---

Hypothesis

We performed a Welch's Two-Sample T-Test to verify the impact of weather.

• $H_0$ (Null): Rain has no significant impact on hourly demand.
• $H_1$ (Alternate): Rain significantly lowers hourly demand.

---

Tools & Technologies

• Python (Pandas, NumPy)
• Stats (T-Tests)
• Visualization (Seaborn, Matplotlib)
• Dataset Source: London Bike Sharing Dataset (Kaggle)

---

Key Findings

1. The "Rain Tax" is 38.7%

Analysis of 17,414 hourly records shows a statistically significant drop in demand ($p < 0.001$). However, the drop is smaller than anecdotal estimates.

Condition	Avg Demand (Rentals/Hr)	Impact
Clear	1,162	Baseline
Rain	712	-38.7%

2. Weekends take the biggest hit

Commuters are resilient; leisure riders are not.

• Weekdays: Demand drops by 36.1% (Commuters still travel).
• Weekends: Demand drops by 45.4% (Leisure riders stay home).

---

Recommendations

Based on the data, the Operations team should move from static shifts to Dynamic Rostering:

1. Weekdays: Reduce staff by 35% when rain is forecast.
2. Weekends: Reduce staff by 45% when rain is forecast.

---

Statistical Validation

Test Used: Welch's Two-Sample T-Test

• T-Statistic: 20.14
• P-Value: < 0.00001
• Decision: Reject H₀ (Rain significantly impacts demand)

Why Welch's Test? Unlike a standard t-test, Welch's doesn't assume equal variance between groups; critical when comparing Clear (N=6,150) vs Rain (N=2,155) samples with different distributions.

Data License & Attribution

This project uses the London Bike Sharing Dataset, provided by Transport for London (TfL).

Attribution:

• Powered by TfL Open Data
• Contains OS data © Crown copyright and database rights 2016
• Geomni UK Map data © and database rights [2019]

The data is used under the Open Government Licence v2.0.