NZ Police victimisation data (Feb 2022 – Apr 2026). Cleaned from UTF-16 CSV, joined to Stats NZ 2017 Area Unit and 2018 Meshblock geographic boundaries (99.4% and 81.3% join rates respectively).
A 500m x 500m regular grid overlaid on the Auckland urban area. Crime counts are allocated to cells via meshblock centroids, producing a dense spatiotemporal tensor for the models.
ConvLSTM
PRIMARYTrained with MSE on log1p-transformed counts. Powers the live forecasts.
ST-ResNet
SECONDARYCloseness (last 3 months) + Period (same month last 3 years).
| Model | RMSE | Theft | Burglary | Assault |
|---|---|---|---|---|
| Last Month (baseline) | 0.957 | 2.07 | 0.80 | 0.67 |
| Historical Average | 1.295 | 3.05 | 0.66 | 0.52 |
| ConvLSTM | 1.030 | 2.34 | 0.65 | 0.64 |
| ST-ResNet | 1.000 | 2.26 | 0.66 | 0.61 |
Auckland-only grid — other cities use TA-level aggregation
Monthly granularity — finer temporal resolution could capture day/hour patterns
No external features (weather, events, demographics) included
Crime forecasting deserves scepticism. The research literature documents real harms from “predictive policing” — feedback loops where patrolling a flagged area generates more recorded crime there, which then reinforces the model (Lum & Isaac, To predict and serve?, 2016; Ensign et al., Runaway Feedback Loops in Predictive Policing, 2018). This project is deliberately scoped to stay clear of that.
It forecasts counts for 500m grid cells. It never identifies, profiles, or predicts individuals.
A public civic-data project. It is not built for, supplied to, or used by Police to direct patrols or target anyone.
It learns from victimisation reports (crime people say happened to them), not arrest or patrol data, and it drives no enforcement — so it can’t create a patrol-then-confirm loop.
Read it as broad clusters and trends with genuine uncertainty, not a promise about any street. Accuracy degrades the further ahead it forecasts.
Built only on data NZ Police already publish openly (meshblock level, CC BY 4.0), then aggregated up to 500m cells and suburbs. It exposes nothing finer than the source and respects Police suppression of sensitive low-count locations.
Recorded victimisation undercounts unreported crime and reflects reporting behaviour, which varies by area and offence type. The model inherits those biases.