Predicting Death Claim Seasonality in a Life Insurance Company using Historical Death Claims Data and Population Deaths Data

A comparison of the effectiveness of a range of Python-developed time series models in predicting the seasonality of death claim notifications for a life insurance company in Ireland. The time series models included both classical and machine learning models.  Each of the time series models were trained on the company’s historic claims data and on Irish population deaths data. The models assessed include a simple historical average model, SARIMA, Random Forest Regressor, Meta Prophet and an LSTM Deep Learning model. Model performance was assessed against a baseline model which predicted no seasonal variation in death claim notifications over a calendar year. The models were trained on both the life insurance company’s historic death claims data and on Irish population deaths data for the 25-74 age cohort over the same observation periods. The results demonstrated that the machine learning models  trained on population deaths data were generally more effective than the baseline model in forecasting death claim seasonality.  However, the simple historical average model trained on population deaths data was also one of the best performing models and its predictions mirrored the Random Forest Regressor output. In general, models trained on the company's historic claims data underperformed the baseline model.