This site hosts a trained model that predicts the likelihood of a fatality given a shark attack anywhere in the world. It also provides some interesting charts and visualizations derived from the underlying data. The model is currently up to date as of February 2026.  The model is an LGBM model :
LGBM stands for Light Gradient Boosting Machine. It’s a tree-based classification algorithm that returns a model designed to be memory efficient and quick to train.
It has been trained on all recorded shark attacks since these records were actually kept as such is composed of around 6800 records which is not a particularly large dataset, and is also the reason LGBM makes sense for this use case. Since it is essentially a Binary Classifier, I built and tested 3 distinct pipelines; Logistic Regression, XGboost, and LGBM. The model that gave the best accuracy on the test data was eventually used for this task. 
The X row used to predict Y, the Fatality is composed of numeric columns, and text columns that are vectorized using SPACY. The text base Activity description column can support sentences like ("Snorkeling and spearfishing in the lagoon"). Attributes such as Longitude, Latitude, location proximity to previous attacks, the shark species in that area, water temperatures, the Sum of prior fatal shark attacks within 200 km over the previous 10 years, weighted by an exponential recency decay w=e−Δt/τw=e−Δt/τ with τ=4 years (floor = 0.1) have been feature engineered and fed to the model in effort to improve its prediction accuracy on the test data.  The model was then optimized by hyper parameter tuning it using random and grid search algorithms to squeeze the last few drops of accuracy. In the end what I got was a model that manages a 76% accuracy on the test data and an 85% accuracy on the train and full data. Not bad, be it slightly overfitted, its fun to use and informative.  

The data used for this model is available at GLOBAL SHARK ATTACK FILE an informative site backed by various shark research societies composed of scientists and journalists. They maintain and update the incident file and I am grateful to them.