BIXI Montréal — Demand Prediction Report 2018

01

Executive Summary

Key findings at a glance

Total trips analysed

869,820

3 months of OD records

Member ridership

89.8%

vs 10.2% casual riders

Decision Tree R²

0.83

83% variance explained

Peak hour demand

98,951

trips at 17:00 (Oct)

"Demand for BIXI bicycles is strongly governed by two predictable forces: commuting rhythms on weekdays and afternoon leisure on weekends. Adding temperature as a variable brings prediction accuracy to R² = 0.83, enabling proactive rebalancing and staffing decisions."

— Nataliya Abboud, Data Science Consultant

Problem

Demand fluctuations — hourly, weekday vs weekend, and weather-driven — make station rebalancing and bicycle availability planning extremely difficult for BIXI's operations team.

Solution

Merged hourly weather records with aggregated BIXI trip data. Engineered time features and built two ML models. Decision Tree selected as the preferred model based on lower RMSE and higher R².

Best model

DT

Decision Tree · RMSE ≈ 407 · R² ≈ 0.83

Primary KPI

Rides/hr

Target variable: total rides commencing within each calendar hour across all active stations.

02

Business Context

Why demand prediction matters for BIXI

BIXI Montréal operates one of North America's largest public bike-sharing networks. Accurate demand prediction is critical for three operational priorities: reducing bicycle shortages at peak hours, improving station rebalancing efficiency, and optimising operational costs across the network.

BIXI's management identified two distinct demand patterns — weekday and weekend — and believed weather to be a significant driver. This project validates both hypotheses with data and delivers actionable ML-based forecasting.

Challenge 1

Peak-hour shortages

Empty docks and missing bikes during 8am and 5pm commute surges cost BIXI in user satisfaction.

Challenge 2

Inefficient rebalancing

Trucks redistributing bikes without demand forecasts waste fuel and labour, often rebalancing too late.

Challenge 3

Weather uncertainty

Cold snaps and rain dramatically suppress ridership, but without a model, staff changes are reactive.

03

Data & Methodology

From raw records to predictive features

Dataset 1 — BIXI Trip Records

Source	BIXI Montréal OD Files
Months	April, October, November 2018
Total records	869,820 trips
Key fields	start_date, station codes, duration, member flag

Dataset 2 — Montreal Weather

Source	Government of Canada
Station	McTavish Reservoir (McGill)
Frequency	Hourly observations · 5,806 rows
Key fields	Temp (°C), dew point, wind speed

1

Load & concatenate trip files

Combined OD CSV files for all available months into a single dataframe. Removed null records to ensure clean aggregation.

2

Feature engineering

Derived hour of day, date, and is_weekend boolean from the start_date timestamp to capture temporal demand patterns.

3

Aggregate to hourly demand

Counted rides per (date × hour) group to create the target variable — trips — matching the granularity of weather observations.

4

Merge weather data

Joined on shared date and hour keys. Dropped rows with missing temperature values to maintain data integrity for modeling.

5

Train / test split & model training

80/20 random split (seed=42). Trained Linear Regression and Decision Tree (max_depth=5) on identical features: hour, is_weekend, Temp (°C).

04

Key Findings

Demand patterns & weather correlation

Hourly trip volume — all days combined

Total trips per hour across April, October & November 2018

Trip volume

Finding: Two clear demand peaks emerge — a morning commute spike at 08:00 (77k trips) and a stronger evening peak at 17:00 (98,951 trips). Night hours (00:00–05:00) see dramatically suppressed demand, enabling maintenance scheduling.

Weekday vs weekend pattern

Average rides per hour by day type

Weekday

Weekend

Weekdays show a sharp dual-peak commuter pattern. Weekends display a single, flatter leisure peak around 14:00–16:00.

Monthly ridership

Total trips by available month

October 2018 recorded the highest ridership (488,490), likely due to warm early-fall weather. November's rapid drop aligns with temperatures averaging −0.2°C.

Temperature vs monthly ridership — seasonal relationship

Average temperature (°C) and available monthly ridership across 2018

Avg Temp (°C)

Ridership (available months)

Finding: Temperature and ridership move in close alignment. The correlation confirms BIXI's management hypothesis — warmer days yield significantly more rides, providing a reliable weather-based planning signal.

05

Predictive Models

Model comparison & selection

Two supervised learning models were trained on the merged dataset using features hour, is_weekend, and Temp (°C) to predict hourly trip counts. An 80/20 train-test split was applied with random seed 42.

Model performance summary

Model	RMSE	R²	RMSE vs best	Status
Decision Tree max_depth = 5	≈ 407	0.83	55%	Selected model
Linear Regression Ordinary least squares	≈ 740	0.44	100%	Baseline

RMSE comparison

Lower is better

R² comparison

Higher is better

"The Decision Tree model captures non-linear relationships — demand does not scale linearly with temperature or hour. At 24°C versus 20°C, the marginal lift in ridership is greater than at 5°C versus 1°C. This threshold behaviour is what Linear Regression misses and Decision Tree captures."

— Model evaluation rationale

06

Recommendations

Operational & strategic actions for BIXI

01

Commute-hour pre-positioning

Increase bicycle availability and docking capacity at major stations before the 07:30 and 16:30 windows on weekdays. Decision Tree forecasts an average of 77k+ trips at 08:00 — staffing rebalancing trucks before these windows reduces shortages.

Trigger: Weekday 07:30 & 16:30 windows

02

Temperature-responsive staffing

Assign additional rebalancing staff on days forecast above 20°C. The model shows strong demand uplift with temperature; warm days (+20°C) drive significantly more trips than cool days (~7°C, as seen in October).

Threshold: Temp ≥ 20°C → elevated staffing

03

Weekend midday rebalancing

Shift weekend operations from a morning-focused schedule to a midday window (12:00–16:00). Weekend patterns show leisure-driven afternoon peaks rather than commute spikes, meaning early morning rebalancing is wasted effort.

Weekend peak: 14:00–16:00

04

Deploy Decision Tree for forecasting

Adopt the Decision Tree model (RMSE ≈ 407, R² = 0.83) as BIXI's short-term demand forecasting engine. Integrate next-day temperature forecasts and day-of-week flags to generate hourly demand predictions for each station cluster.

Model accuracy: R² = 0.83

Data sources

BIXI Montréal (2018)

BIXI trip data — origin-destination records for April, October and November 2018.

Environment Canada (2018)

Historical hourly weather data — McTavish Reservoir station, Montréal.

McGill School of Continuing Studies

Case study instructions and analytical framework.

Analysis tools

Python (pandas, scikit-learn, matplotlib, seaborn) · Alteryx · Tableau

Predicting Bicycle Demand for BIXI Montréal