Brought to you by:

Machine learning shows promise for hazard budgeting

Facebook Twitter LinkedIn Google

There are promising results in applying machine learning techniques to the significant challenge of accurately predicting natural climate variability, attendees at this year’s Actuaries Summit heard.

The ability to predict the cycles of El Niño-Southern Oscillation (ENSO), Indian Ocean Dipole (IOD) and Southern Annular Mode (SAM) – key drivers for extreme weather events – is currently limited beyond 6-12 months.

This limitation has contributed to disclosure requirements “leap-frogging” the capabilities of climate science by at least a decade, Suncorp Natural Perils Senior Pricing Advisor Tatiana Potemina says, noting this natural variability is a key element of uncertainty in climate risk modelling.

Extending the forecasting range will be vital in helping insurers manage pricing, reinsurance design and budgeting and experts are hopeful technology may hold the solution.

“Improvements so far have been incremental rather than a step change but machine learning and neural networks are increasingly being used and the results are really promising,” Ms Potemina says.

“They can predict to 17 months in advance, so that’s quite significant,” she said. “There are some promising results there.

“For insurers, it’s important to understand natural climate variability because without this understanding some incorrect assumptions can be made about the trend.”

Ms Potemina points to a scientific paper called “Deep learning for multi-year ENSO forecasts” which notes that while robust, long-lead forecasts would be valuable for managing policy responses, decades of effort has not resulted in lead times of more than one year for forecasting ENSO events.

“We show that a statistical forecast model employing a deep-learning approach produces skilful ENSO forecasts for lead times of up to one-and-a-half years,” the paper says.

That study used transfer learning to train a convolutional neural network (CNN) on historical simulations and reanalysis. It found the correlation skill of the CNN model was “much higher than those of current state-of-the-art dynamical forecast systems”.

“The CNN model is also better at predicting the detailed zonal distribution of sea surface temperatures, overcoming a weakness of dynamical forecast models,” the paper says.

“The CNN model is a powerful tool for both the prediction of ENSO events and for the analysis of their associated complex mechanisms.”