Highest ever monthly number recorded; 5x higher than the usual yearly peak
Resource Planning Forecast: NYC Pothole Maintenance Demand
Purpose: Enable the Department of Transportation (DOT) to proactively assemble crews, materials, and budget to handle fluctuating and recurring pothole maintenance
Result: A 9-month forecast of monthly NYC 311 pothole complaints (Aug 2026 – Apr 2027)
Background: Following a period of unusually mild winters (2020–2025), the record snowfall of 2026 has significantly shifted demand patterns. To manage this new volatility, the DOT must transition to more robust, outlier-resilient forecasting to ensure resource readiness.
Notable Pothole Complaint Numbers
Typical summer demand but definitely an underestimate
Expecting ~2x more than typically seen from 2020-2025
Understanding NYC's Pothole Cycles (2009–2026)
Winter snowfall leads to water seeping into asphalt and freezing. As winter ends in February/March, the ice melts leaving cavities in the asphalt and causing potholes. This causes the yearly complaint peak in March.
NYC 311 Monthly Pothole Complaints History (Dec 2009 – Jul 2026)
2015 Winter Thaw
- Shifted baseline complaint numbers upward by ~20%
- Large winter events have multi-year lasting effects on road conditions
2026 Winter Thaw
- All-time record of 25,171 complaints
- 5.1× baseline extreme freeze-thaw anomaly
- Expect significantly elevated complaint volumes going forward
9-Month Resource Demand Forecast (Aug 2026 – Apr 2027)
Complaints will surge to 11k in March 2027 (~4x the autumn baseline). The forecast is presented alongside monthly data from the last major snow year, 2015, to provide context.
Monthly Complaint Projections with 95% Confidence Bands
Analyst Recommendation for Leadership
Action Required: Given 2026's record snowfall, complaint volumes will likely exceed baseline forecasts for the next few months as the model adjusts to the anomalous data. Operations should scale capacity closer to the upper forecast range.
Why Winter Matters: The Freeze-Thaw Multiplier Effect
Monthly Seasonal Multiplier (How Winter Amplifies Demand)
0 = Baseline. Positive values multiply complaints above baseline; negative values reduce them.
March: Peak Demand Period (+148%)
- Annual peak in all 16 historical years
- This predictable pattern enables accurate forecasting
July-December: Minimal Complaints
- Annual complaint trough across NYC
- Dry, warm summer and autumn weather minimizes pothole formations
Methodology: Multiplicative Scaling
- Surges scale proportionally with baseline (+50% to +150%)
- Multiplicative model prevents summer overprediction
Stress Test: How the Model Handled Outlier Extreme Weather in its Test Dataset
A model was selected based on its cross-validation performance from 2009 to July 2025. The model was then tested on August 2025–July 2026 data. March 2026 had a record: 25,171 actual complaints vs. the predicted 6,807. While this outlier winter event tested the model's limits, the forecast maintained a 72.8% accuracy (27.2% error).
Actual vs. model predictions. (Aug 2025 – Jul 2026)
Prophet Model Hyperparameter Selection: Validation Matrix
Compared 32 different models across 4 trend model configurations (Additive Smooth, Additive Flex, Multiplicative Smooth, Multiplicative Flex) and 8 different changepoint parameters. Multiplicative seasonality with explicit event changepoints achieves the best cross-validation generalization score.
Additive Smooth
| Hyperparameter | CV | Test |
|---|---|---|
| Events | 25.9% | 33.5% |
| Flex | 26.5% | 33.5% |
| Auto | 26.6% | 36.2% |
| None | 26.9% | 37.6% |
| Evt+Reg | 35.9% | 25.6% |
| Reg | 40.1% | 27.4% |
| Auto+Reg | 41.8% | 23.3% |
| Flx+Reg | 49.2% | 27.2% |
Multiplicative
| Hyperparameter | CV | Test |
|---|---|---|
| Events | 18.7% ★ | 27.2% |
| Auto | 19.7% | 32.4% |
| None | 21.7% | 33.9% |
| Flex | 22.5% | 32.4% |
| Reg | 31.3% | 28.2% |
| Evt+Reg | 31.6% | 24.5% |
| Flx+Reg | 32.2% | 24.4% |
| Auto+Reg | 32.3% | 26.2% |
Additive Flex
| Hyperparameter | CV | Test |
|---|---|---|
| None | 26.9% | 37.5% |
| Auto+Reg | 32.3% | 39.6% |
| Flx+Reg | 32.7% | 39.7% |
| Evt+Reg | 33.2% | 37.3% |
| Auto | 33.3% | 50.3% |
| Events | 33.6% | 51.4% |
| Flex | 34.0% | 50.8% |
| Reg | 40.1% | 27.4% |
Flat Trend
| Hyperparameter | CV | Test |
|---|---|---|
| None | 34.9% | 28.6% |
| Auto | 34.9% | 28.6% |
| Events | 34.9% | 28.6% |
| Flex | 34.9% | 28.6% |
| Reg | 43.4% | 30.1% |
| Evt+Reg | 43.4% | 30.1% |
| Auto+Reg | 43.4% | 30.1% |
| Flx+Reg | 43.4% | 30.1% |
14-Fold Cross-Validation (CV) Data (Dec 2009 – Jul 2025)
- Selected Model (Multiplicative | Events) provides the optimal model parameters
- Adding regressors inflates CV error >31% due to overfitting
Test Data with Extreme Winter Shock Outlier (Aug 2025 – Jul 2026)
- Regressors help fit extreme shocks (24.4% test MAPE beats the selected model's 27.2%)
- The regressors have deleterious recency bias with a MAPE of 32.2% for the 14 years of cross-validation data
Methodology: Data Pipeline, Model Engine & Validation
Data Cleaning & Preparation
- Source API: NYC 311 Service Data Requests via Socrata API (2020-2026 and 2009-2020 datasets)
- Filtering & Matching: Filtered for descriptor
%POTHOLE%and'Rough, Pitted or Cracked Roads'(~918K total complaints aggregated into monthly series) - Outlier & Boundary Handling: Dropped incomplete month (Aug 2026); retained true weather shocks (e.g., Mar 2026 25.2k spike) to calibrate risk without hiding climate volatility
Prophet Model Hyperparameters
- Formulation: Multiplicative Generalized Additive Model (GAM):
y(t) = trend(t) × (1 + seasonality(t)) + events(t) + noise - Hyperparameters:
changepoint_prior_scale = 0.01(prevents trend overfitting);seasonality_prior_scale = 100.0(captures sharp seasonal surges) - Inference Engine: Powered by Stan with Bayesian MCMC sampling (300 draws) for true 95% confidence intervals
Feature Engineering & Regressors
- Seasonality Order: Multiplicative Fourier series (Order N=10) capturing the annual March peak without summer overprediction
- Event Regressors: Binary dummy indicators for known weather shock windows (e.g.,
snowstorm_2015 = 1for Jan–Mar 2015) to isolate temporary anomalies - Trend Changepoints: Automatic changepoint detection combined with explicit event breaks for major municipal shocks
Validation Protocol & Metrics
- MAPE Metric: Mean Absolute Percentage Error:
MAPE = Mean(|Actual − Forecast| / Actual) × 100%(18.7% MAPE = 81.3% accuracy) - 14-Fold Cross-Validation: 81.3% accuracy (18.7% MAPE) across all years 2009–2025
- Held-Out Outlier Test: 72.8% accuracy (27.2% MAPE) on 12 unseen months including the record March 2026 shock
Next Steps
While the Prophet model provides a strong forecasting baseline, there could be further model improvements with supplemental datasets and other model architectures.
New Data Inputs
- Weather: Snowfall, rainfall, and freeze-thaw cycles are the main drivers of pothole formation. Therefore, using temperature, precipitation, and snow-depth data as inputs for the forecast as leading indicators would probably improve the model results
- Cave-in complaints: "Cave-ins" are potholes formed due to freeze-thaw-induced cavities under the asphalt instead of within the asphalt. While "cave-in" complaints are handled by a different department, there are ~168K cave-in records in the 311 dataset and including them could provide more information to improve forecasting
Alternative Model Architectures
- Gradient-boosted trees (XGBoost, LightGBM, CatBoost) can capture non-linear interactions between weather and complaint history that an additive GAM cannot express
- Foundation forecasters (AWS Chronos, Google TimesFM) treat time-series prediction as a language-modeling problem
- Reinforcement learning could optimize the sequential crew-allocation and resource-deployment problem under evolving forecast uncertainty
- Diffusion models are emerging for extreme-event prediction: they model complex, multi-modal uncertainty distributions, making them well-suited for rare high-variance shocks like the March 2026 spike
Applicant: Christian Rivera
Background & Profile
- Specialist in quantitative forecasting, investment portfolio construction, and software development
- My cats are named after U.S. states
- My daughters are named after French mathematicians
Technical Stack For This Project
- Python for data analysis & forecasting
- HTML5 & CSS3 for structure & design
- Vanilla JavaScript for interactivity
- Cloudflare Pages for static site hosting
- The Vanguard style guide
See Some of My Other Work
- Github Pages Pages Portfolio
- Fixed Income Portfolio Signal Attribution Analysis (upon request)
- Project Management Reporting Dashboard (upon request)
AI Tool Assistance Attribution
- Gemma4-26B-A4B local model
- Qwen3.6-27B local model
- Pi Agent coding harness
- Local AI work is a hobby of mine. Ask me about it.