The single strongest predictor. Above 45mm in 48 hours, murky runoff flushes baitfish to the surface and sharks investigate by touch rather than sight — significantly increasing bite risk. Above 100mm, risk is extreme. The January 2026 Manly cluster followed 135mm in 24 hours.

Warmer water = higher bull shark risk. Activity peaks at 23°C and becomes meaningful above 20°C. Below 18°C, bull sharks largely vacate Sydney-latitude beaches. Note: this is the opposite to white sharks.

December through March is peak season. Risk remains elevated through November-May as water stays warm. Season captures the structural annual pattern that temperature alone can miss at shoulder months.

1.8–2.8m is the sweet spot for bull shark inshore hunting. Flat conditions are too exposed. Very large swell (above 3.5m) pushes sharks offshore. Swell amplifies other factors but won't drive extreme risk alone.

Risk increases sharply within 1km of a river or lagoon mouth. Bull sharks are euryhaline — they move freely between salt and fresh water. For ICOLL lagoons (Dee Why, Avoca), lagoon state is estimated from recent rainfall.

Bull sharks do not follow whale migration corridors. This factor is only included in the white shark model.

The strongest single driver. White sharks follow humpback whales — they target cow-calf pairs as prey. The southward migration (August-November) passes directly through NSW waters, peaking mid-September. This happens every year, regardless of other conditions.

Cooler water = higher white shark risk — the opposite of bull sharks. Attack sites average 19.7°C. White sharks seek cool upwelling pockets to ambush prey with a visual advantage. The September 2025 Long Reef fatal attack occurred in 19.5°C water.

Spring (September-November) is peak, aligning with the southward juvenile migration. This peak has shifted about three weeks earlier and is now centred on mid-September. A secondary elevation occurs in winter (June-August) during the northward run.

Risk elevated within 10km of a river mouth — a much wider radius than bull sharks. This reflects prey concentration near outflows across a broader area, not a preference for estuaries.

Rainfall concentrates prey near shore, providing a secondary signal. It does not directly drive white shark behaviour the way it does for bull sharks. Significantly less important than for bull sharks.

Moderate swell (1.0-2.5m) provides marginally better hunting conditions. Swell alone has little predictive power for white shark presence. The lowest-weight factor in the model.

The factors in this model — rainfall, water temperature, season, swell, river proximity, and whale migration — are drawn from peer-reviewed research, primarily Smoothey et al. (2023) on bull shark occurrence in NSW coastal waters and Ryan et al. (2019) on predictive modelling of shark attacks in Australian waters. Both studies used generalised additive mixed models (GAMMs) to identify which environmental variables significantly predict shark presence or attack risk. The threshold values (e.g. 45mm of rain, 22°C for bull sharks) reflect the breakpoints identified in that research.

The published studies identify which variables matter and their direction, but do not produce percentage weights that can be directly applied to a scoring model. The relative weights in SharkCast — for example, rainfall carrying more weight than swell for bull sharks — were hand-set to reflect the ordering of importance described in the research literature. They are research-informed but not statistically derived. The honest characterisation is that the factors and thresholds have a scientific basis; the precise weights are calibrated judgement.

Rainfall decay assumption: The model applies a 3-day weighted window to rainfall — 100% weighting for the day of rain, 100% for the following day, and 40% for two days after. This reflects the physical dynamics of NSW coastal ICOLLs, where turbidity typically peaks 12–24 hours after a rainfall event and returns to baseline over 48–72 hours. These weightings are an informed assumption based on estuarine flushing dynamics, not a value published in the shark research literature. Empirical turbidity monitoring data from specific lagoons (e.g. Dee Why, Narrabeen) would allow these to be refined.

The model was tested against documented NSW shark attacks with known dates and locations, covering both species and a range of conditions. Where environmental data could be reconstructed — rainfall totals, sea surface temperature, month — conditions were scored and results compared against what actually occurred.

The model performed well at the extremes. Incidents with well-documented extreme conditions scored in the Extreme tier. Incidents that occurred during conditions not strongly associated with shark activity scored in the Low to Moderate range — consistent with the research position that some attacks are not predictable from environmental inputs alone.

One key finding from the back-test: the model's primary value is at the tails. It reliably flags genuinely dangerous conditions — the kind that produced the January 2026 Sydney cluster — and reliably identifies genuinely low-risk days. In the middle range, it provides a signal but not a prediction.

Some attacks occur on days that score Low or Moderate. This is not a model failure — it reflects a genuine limit of environmental prediction. A single shark being in the wrong place at the wrong time, the presence of dolphins or baitfish in the immediate area, or individual animal behaviour cannot be captured by date, temperature, and rainfall data. Low risk does not mean no sharks. It means conditions do not systematically favour nearshore shark activity. SharkCast is a risk indicator, not a safety guarantee.

Sea surface temperature, rainfall and swell are fetched from Open-Meteo's weather and marine APIs daily. SST is now forecasted per day using the daily marine API — the temperature line chart shows a real 6-day trend, not a flat monthly average.

Tagged shark locations are the single biggest gap. NSW DPI's SharkSmart program tags hundreds of sharks but doesn't publish a real-time API. A confirmed tagged shark near your beach would override everything else in the model.