How Does AI Predict When Fish Will Bite?

Two different things get called AI

Most fishing apps that advertise AI are running a rule-based score rather than a learned model, and the distinction matters more than the marketing suggests.

A rule-based score works the way you would expect if you sat down and wrote it yourself. Rising pressure adds points. A stiff wind adds a few. A cold snap takes some away. Someone decided each of those weights, and the result is a number you can take apart: every factor that moved it can be listed. Our own BiteScore is built this way, and we say so on its page.

A trained model is a genuinely different animal. It is shown a large set of recorded outcomes — what the conditions were, and what actually happened — and it works out for itself which patterns held. The crucial consequence is that a trained model can be measured: you hold back some data, ask the model to predict it, and count how often it was right.

That gives you a clean test for any product claiming AI. Ask what it was measured against. A trained model has an answer. A rule-based score does not have one, because there was never a training set to hold back — and that is not damning, but it does mean an accuracy percentage would be invented.

What the evidence says conditions actually do

Environmental conditions do measurably affect catch rates, and the effects are consistently smaller than fishing content implies.

The largest study we could find analysed 341,959 angler-reported muskellunge catches from 1970 to 2013. It found catch was statistically related to the 29-day lunar cycle — and put the maximum overall effect at about 5%. In the authors’ framing, someone fishing only on the peak lunar day would average around 5% more fish than someone fishing random days. Catch never approached zero during the worst lunar periods.

Context changed the size of it. The effect rose to roughly 28% for night fishing, and to about 26% at one particular water body. That variation is itself informative: this is not one universal rule but something that depends on species, water and time of day.

Temperature has firmer physiological grounding. Most fish are ectotherms, and a review of temperature effects on feeding describes how it governs metabolic rate, energy balance, the ability and the desire to obtain food, digestion and nutrient absorption. But the same review is explicit that these effects are complex and species-specific, varying with the timing, intensity, duration and rate of the temperature change. There is no single curve that applies to every fish.

Oxygen matters for reasons that are well documented without being about bite timing. The US Geological Survey notes that cold water holds more dissolved oxygen than warm water, and that levels below 2 milligrams per litre are considered hypoxic. The Environmental Protection Agency adds a detail anglers can use: dissolved oxygen tends to be lowest just before dawn, and fish kills characteristically occur between about two in the morning and sunrise.

The trap hiding inside catch data

Catch records measure two things at once — when fish fed, and when people went fishing — and separating them is harder than it sounds.

The muskellunge study is admirably direct about this. Angler effort itself concentrated around the full and new moons at one of the lakes studied, which meant the authors could not conclude that the lunar pattern came from fish behaviour alone. More fish were reported caught partly because more anglers were out.

Follow that through and it has an uncomfortable implication for any forecast trained on catch reports. A model learning from that data learns when people fish as much as when fish feed. It then recommends those same times, more anglers go out, more catches get reported, and the pattern confirms itself. The prediction is partly manufacturing its own evidence.

There is no clever way around this using catch reports alone. It takes a study design that measures effort separately — which is precisely why telemetry work, tracking fish directly rather than counting what anglers caught, is more trustworthy on this question.

What happens when the science tries it properly

Machine learning does real work in fisheries science, and the published accuracy figures range from excellent to worthless depending on how much data there is.

At the strong end, one study trained random forest and gradient boosting models on roughly 10,000 labelled fishing trips and classified whether a vessel had fished in state or federal waters at about 97% accuracy. That is a genuine result — but notice what the task was. It classified records of things that had already happened, using a narrow, well-defined question with abundant labelled examples. It did not predict fish behaviour.

The other end is more instructive. A bycatch-estimation study using ensemble random forests reported test performance ranging from 0.97 for Laysan albatross down to 0.48 for leatherback turtles — and 0.48 is no better than a coin toss. Same method, same team, same ocean. The difference was data: for rare species there simply were not enough records to identify environmental correlates at all.

There is a further warning worth carrying. Work on ecological inference from machine learning found that including variables which correlate with the outcome but have no causal relationship to it can actively interfere with drawing conclusions — and that increasing the sample size does not fix the problem. More data does not automatically make a bite predictor valid.

The measurement nobody has published

We could not find a single peer-reviewed evaluation of consumer bite-forecast accuracy, and we think that absence deserves stating plainly rather than being glossed over.

The nearest thing is research testing the underlying idea. The telemetry study of largemouth bass mentioned above tracked 22 adult fish through a whole year with a lake-wide acoustic array. Lunar variables did affect swimming activity and depth — but not consistently across seasons, and the authors concluded that solunar tables consulted by anglers “may have little predictive value for identifying peak fishing time.” They also noted the seasonal effect dwarfed the lunar one: daily movement distances were around five times greater in spring and summer than in winter.

So the position an honest reader should hold is this. Some inputs these forecasts use have solid physiological grounding, particularly temperature and oxygen. One widely marketed input — the moon — has been tested directly and came out weak. And the forecasts themselves, as products, have not been independently measured by anyone we can cite.

That applies to ours. We would rather tell you that than quote a number we made up.