Liv-ex’s ‘Fair Value’ methodology

Many Liv-ex members will be familiar with the Fair Value Tool, which we have used in our En Primeur analyses to assess the pricing of new releases. Liv-ex data scientists, through a thorough review of the methodology, have determined that log-linear rather than polynomial regression better represents the relationship between Market Prices and scores, providing members with more reliable insights.

Over the years, the relationship between critic scores and prices has changed. In the past, critic scores – particularly Robert Parker’s — were strong predictors of price. With more critics entering the scene and technological innovations in the vineyard and cellar allowing for greater homogeneity between vintages, the relationship has broken down.

 While the previous iteration of the Fair Value tool has provided buyers with useful insight over the years, the time has come to review and refresh the methodology.

Polynomial regression (the previous methodology) worked well in capturing the exponential relationship between Robert Parker scores and Market Prices. With a small data set (less than 20 vintages) and the relationship having broken down, however, we started to see instances of the regression line overfitting the data. This effect can result in non-sensical interpretations of Fair Value lines.

Consider, for example, Market Prices of Château Haut-Batailley vs. Neal Martin scores, using polynomial regression. While polynomial regression has the benefit of capturing increasingly steep growth in price – the price difference between 99 points and 100 points is often much wider than 98 and 99 – it can also fit the data too closely, creating an unrealistic U-shaped curve (negative reciprocal). In this example, it appears that the Fair Value of a 91-point and 95-point wine are equal, while 92-, 93-, and 94-point vintages should be less expensive.

Considering the wines we will cover during this year’s En Primeur campaign (which includes all comprising the Bordeaux 500) vs. Neal Martin scores, Liv-ex data scientists tested five different statistical models: linear, log-linear, polynomial, Poisson and spline. Each method was evaluated using the AIC and BIC methods, which, in essence, measure the average distance from regression lines formed to the actual data points (read more here). Since AIC and BIC measure distance from actual data points, the lower the score, the better. The analysis revealed that the log-linear model had the lowest AIC and BIC scores, making it the best fit for the data.

Considering Château Haut-Batailley again, log-linear regression returns a near 0 R-squared. In this case, this is a better representation of the data – the line has not been overfit. A lack of correlation here is informative – there are other factors, vintage reputation for example, that are driving differences in price.

The new methodology can be used in much the same way as before, but with better reliability. Wines that fall below the line represent Fair Value insofar as they are less expensive than the theoretical average for a wine of that age or rating. The higher the R-squared, the more of the variance in price is explained by age or rating. In the above example, a potential buyer would be correct in understanding that: