To overcome liquidity problems responding to disasters, the World Bank supported the design and implementation of a flood index insurance for Myanmar, Lao PDR and Cambodia. The policy itself is offered by an Insurance Facility (IF) in Singapore, capitalised by international donors. The index insurance uses a hydrodynamic flood model to base its premium and payout on. However, the hydrodynamic model contains falses (both positive and negative) and to validate the model, the Insurance Facility seeks from-the-ground observations. Especially the negative falses have high priority.
To provide from-the-ground observations and validate the model, FloodTags developed an operational system that detects flood events from online news and social media in the respective countries. Using machine-learning we trained native-language classifiers, detecting new flood events with overall confidence (F1-score) per article of 82%. As most events are covered in multiple articles, errors are corrected and there are hardly any events missed. With high precision and recall, the system:
Analyses media data to find evidence of a modelled flood: If a flood has been identified by a model, the user can easily see whether the flood has also mentioned in the social or other online media. If this is not the case, the model output could be erroneous and follow-up investigation by an expert is advised.
Informs a user about new floods that may otherwise have gone unnoticed by the model. As soon as the media report on a new flood (large or small) it is sent to the user. If the modelling platform does not show any flood at that moment, follow-up investigation by an expert is advised (to confirm which of the sources has the better information).
The system is now prepared for operations at the Insurance Facility in Singapore.
Index insurances are extremely effective tools in poverty alleviation. Flood, drought and precipitation index insurances gain in popularity worldwide. But there are serious challenges in the policy design and operation, such as:
How to obtain sound historic data for trigger design and premium setting and
How do you accurately monitor the situation in real-time and validate the underlying models (staying aware of false positives and false negatives.
The generally available event databases (such as EM-DAT or DesInventar for floods) are often not complete enough and they come with a delay. Using media data as additional data source is very effective to cope with these challenges. Even in low resourced languages, as we have found with Burmese and Lao.
If you would like to validate your models for index insurance, and you are interested in the possibilities of media monitoring, please email us for an introduction meeting without charge.