The reconstruction value of a house is the total cost to reconstruct that house (in case of total loss due to for example a fire or natural disaster, and therefore not to be confused with the transaction price of a house). Since risk equals probability times consequence, knowing the reconstruction value seems important in order to estimate the risk of a house insurance. Most insurers rely significantly on the estimation of the reconstruction value in order to calculate the appropriate insurance premium.
We started looking into this topic since we noticed that many property insurers are unhappy with the accuracy with which these reconstruction values are calculated. It is not unusual that when a house fully burns down, the actual payout that has to be made by the insurer is very different compared to the reconstruction value that was estimated at the inception of the insurance contract and based on which the insurance premium was calculated. This is particularly annoying when the reconstruction value has been underestimated as this will result in a larger than expected loss for the insurer. If there is indeed this love-hate relationship with reconstruction value, why do insurers hang on to this that much? How important is this reconstruction value estimation really?
A fire insurance covers a lot more than only the complete loss of a house. Today, most Belgian fire insurance policies also cover costs such as:
Through our work with the insurance industry, we have come to the conclusion that, on average, the total of these “smaller” costs (insurance payouts of less than EUR 25k per claim) make up more than 90% of the sum of all insurance payouts.
This also means that for 90% of the claims a home insurer receives, the reconstruction value is not the strongest (and often not even a good) predictor of the expected costs related to these claims. Other, more objective and straightforward, characteristics of the house (e.g. location, volume, roof type, height, …) have a much higher predictive power.
The concept of reconstruction value can of course not be entirely dismissed either. Occasionally, a house will be completely destroyed and the insurer will need to pay out the actual reconstruction value. In that regard, having an accurate estimation of the reconstruction value remains relevant.
In the following paragraphs, we discuss two different ways to estimate this reconstruction value: either with a surveyor or via a statistical model. We claim that both methods have serious flaws.
Insurance companies work with surveyors (also called experts) to evaluate a house and estimate its reconstruction value. Surveyors are typically called upon in two types of situations: they evaluate either a house which has not (yet) incurred any damage; based on this evaluation, an insurance premium can subsequently be determined. Or, they will evaluate the damage that has effectively occurred which in turn will determine the insurance payout.
Humans learn by practice. Whether its a professional chess player, a firefighter or a doctor, they usually always manage to improve their trade through years of experience. The mechanism with which one can gather additional experience very much depends on the amount of time it takes to get feedback about your actions and decisions. A heart surgeon will get immediate feedback when he makes a mistake. The consequence of his mistake could very well result into the death of his patient but he will most likely never make that same mistake again. When a psychiatrist makes a mistake though, which could possibly also have deadly consequences, he may never find out. The same applies for an engineer that designs formula-one racing cars (and learns during tests/races if his upgrades worked) compared to an engineer that designs bridges that are built only years later (and might only fail 20 years later). These concepts are well explained in the book “Thinking Fast and Slow” by Daniel Kahneman.
We believe the same analogy holds for surveyors who estimate the reconstruction value of a house vs. surveyors who estimate the transaction value of a house. In the latter case, a surveyor (or real estate agent) will estimate up to several hundred of houses every year. The majority of these will be sold within a couple of weeks. The actual price at which the house is sold is very valuable feedback for the surveyor since it will allow him to get better at estimating the right transaction price.
Surveyors making estimations of reconstruction values do not benefit from the same feedback mechanism, as only very few of the estimated houses will have to be rebuilt at some point in time. Even in the event that a house is damaged to the point that it requires to be completely reconstructed, it is very likely that multiple years have passed by then. Even if the expert were to be informed of the actual reconstruction value related to a house he historically evaluated, it would probably have very little value since he or she will in all likelihood not remember the very specifics of the estimation. The lack of feedback which is inherent to estimating a reconstruction value makes us question the accuracy of these estimations made by surveyors.
It is quite common today to estimate the market value of a house with a statistical model. Many features of the house, such as the location and volume are well correlated with the transaction value of the house. These correlations can be learned by analyzing real estate ad listings. In some countries the government will even have a public registry with the details of all real estate transactions. Once you understand these correlations, you can come up with some kind of regression model to predict the market value of basically any house.
Can such a regression model also be built to predict the reconstruction value of a house?
Unfortunately, there is no dataset with thousands of actual reconstruction values available. As mentioned earlier, very few houses burn down every year, and even for a larger insurer this results in only a limited number of data points to work with. As opposed to the real estate ad listings, that can be used to train a predictive model for transaction prices, we don’t see an equivalent training data set to develop a regression model for estimating reconstruction values.
What we do not consider as a valid training dataset is a list of estimations of reconstruction values made by surveyors. Using building features to predict these estimations would only result in mimicking their process. In the previous paragraphs we have already pointed out that due to a poor feedback mechanism, this makes little sense.
Nowadays, most insurers try to approximate the reconstruction value of a house by using a statistical model. One of the popular building parameters is the number of rooms a house has.
Old houses in Belgium often have many smaller rooms downstairs (living room, kitchen, storage rooms, bathrooms, …). New and renovated houses on the contrary have very few rooms on the ground floor. Kitchen, dining room, living room, etc. are often all in one large space. The statistical models of many insurers do not really take these new building trends into account though. This can lead to significant differences in estimated reconstruction values for houses that are, from a construction point of view, very similar.
While we do recognize that being “roughly right” is often better than “exactly wrong”, we do believe that a lower weight could be given to the reconstruction value in determining the overall risk of a house insurance.
Instead, more focus should be put on the historical claims data and the correlations they can uncover between the observed claim frequency/severity and certain building characteristics (e.g. volume, height, roof type, construction type, …). It is beyond the scope of this blog post to enter into much more detail on how to do this. However, don’t hesitate to get in touch with us if you want to hear more about this.