Dynamic Pricing Engine for Urban Parking Lots

Forecasts next-day parking demand using XGBoost and sets prices proactively — reducing overcrowding at full lots and increasing utilization at underused ones.

Summer Analytics 2025 · Consulting & Analytics Club × Pathway

---

The Problem

Urban parking lots charge a flat rate all day regardless of demand. This causes two problems simultaneously:

• Peak hours → lots overflow, cars circle the block wasting fuel, drivers frustrated
• Off-peak hours → lots sit half empty, operators lose revenue

The solution is dynamic pricing — automatically raising prices when demand is high and lowering them when it is low. Same logic as Uber surge pricing, applied to parking.

---

What This Project Does

1. Loads 18,368 real parking records from 14 lots in Birmingham, UK (73 days, every 30 minutes)
2. Cleans the data — parses timestamps, encodes categorical columns, computes occupancy rate
3. Processes it through a Pathway streaming pipeline using tumbling window aggregation
4. Applies 3 rule-based pricing models per day (baseline, demand-based, competitive surge)
5. Forecasts next-day peak occupancy using per-lot XGBoost models trained on 16 engineered features
6. Sets prices proactively — so the right price is ready before demand hits, not after
7. Alerts when any lot exceeds 90% occupancy and identifies nearby alternatives using Haversine distance

---

Results

Metric	Value
Best lot R²	0.77
Average MAPE across all lots	9.56%
Lots with reliable ML predictions (R² > 0.4)	5 of 14
Rerouting alerts generated	37
Price range (clamped)	$5 — $20

---

Architecture

Raw CSV (18,368 rows · 14 lots · 73 days)
        │
        ▼
Data Cleaning (Pandas)
  - Parse timestamps
  - Encode traffic / vehicle type categories
  - Compute occupancy rate
        │
        ▼
Pathway Streaming Pipeline
  - Processes data row-by-row
  - Tumbling window aggregation (daily)
  - 3 pricing models run in parallel
        │
        ├──────────────────────────────────┐
        ▼                                  ▼
Rule-Based Prices                  XGBoost Forecasting
(all 14 lots)                      (one model per lot)
                                   - 16 engineered features
                                   - Lag variables
                                   - 7-day rolling averages
                                   - Time signals
                                          │
                                          ▼
                                   Hybrid Pricing System
                                   - ML pricing: 5 reliable lots (R²>0.4)
                                   - Rule fallback: 9 erratic lots
        │
        ▼
6-Panel Matplotlib Dashboard
+ Interactive Bokeh Visualizations

---

Pricing Models

Model 1 — Baseline Linear

Price increases with daily occupancy swing (max minus min occupancy that day).

Price = $10 + 5 × (occ_max - occ_min) / capacity
Range: $11 – $15

Model 2 — Demand-Based

Combines 5 signals into a weighted demand score.

Demand = 0.5×(Occ/Cap) + 0.3×QueueLength + 0.4×IsSpecialDay + VehicleWeight + TrafficWeight
Price  = $10 × (1 + 0.8 × Demand / 5)
Range: $15 – $22

Model 3 — Surge / Discount

Clear thresholds with a hard price clamp.

Occupancy > 80% → 1.5x surge multiplier
Occupancy < 30% → 0.8x discount multiplier
Final Price = clamp($5, computed_price, $20)

Model 4 — Competitive

Processes all 14 lots simultaneously. Lots above 85% raise prices slightly; lots below 30% undercut the market. Rerouting alerts trigger at 90% occupancy using Haversine distance to find nearby alternatives within 1km.

---

XGBoost Demand Forecasting

Why per-lot models?

One model trained on all 14 lots combined gave R²=0.45 — each lot has different capacity, location, and demand patterns that confuse a shared model. Training one model per lot gave R²=0.77 on the best lot.

Features (16 total)

Category	Features
Today's data	peak occupancy, avg occupancy, min occupancy, queue max, queue avg, is_special_day, traffic avg, vehicle avg
Time signals	day of week, month, week of year, occupancy rate
Lag features	yesterday's peak, yesterday's queue, yesterday's occupancy rate
Rolling window	7-day rolling average (61% feature importance)

Hybrid System Design

Condition	Approach
R² > 0.40 — 5 lots	XGBoost predicted occupancy drives tomorrow's price
R² ≤ 0.40 — 9 lots	Rule-based fallback using actual occupancy

Negative R² means the model performs worse than simply predicting the mean. Forcing ML on unpredictable lots makes pricing worse, not better — so rule-based logic takes over for those.

---

Tech Stack

Tool	Purpose
Python	Core language
Pathway	Streaming pipeline, tumbling window aggregation
XGBoost	Per-lot demand forecasting
Scikit-learn	Train/test split, RMSE, R², MAE, MAPE
Pandas	Data loading, cleaning, feature engineering
NumPy	Numerical operations
Matplotlib	6-panel static dashboard
Bokeh	Interactive visualizations
Google Colab	Development environment

---

Dataset

Real parking occupancy data from Birmingham, UK.

Field	Description
SystemCodeNumber	Unique lot ID (e.g. BHMBCCMKT01)
Occupancy	Cars currently parked
Capacity	Maximum lot capacity
QueueLength	Cars waiting to enter
TrafficConditionNearby	high / average / low
IsSpecialDay	Holiday or event flag (1/0)
VehicleType	car / bike / truck / cycle
Latitude / Longitude	GPS coordinates
LastUpdatedDate / Time	Observation timestamp

• 73 days of data
• 18,368 rows total
• 18 readings per day per lot (every 30 min, 8am–4:30pm)
• 14 parking lots

How to Run

# Clone the repo
git clone https://github.com/yourusername/dynamic-parking-pricing
cd dynamic-parking-pricing

# Install dependencies
pip install pathway xgboost scikit-learn pandas numpy matplotlib bokeh

# Open in Google Colab or run locally
jupyter notebook parking_pricing.ipynb

---

---

Built as part of Summer Analytics 2025 · Consulting & Analytics Club × Pathway