Technical appendix

We use the appendix to provide some additional details concerning our estimates. It is intended to help with both the interpretability and replicability of our results.

Estimating production levels

Bloomberg provides revenue breakdowns by its industry classification system BICS. Note that metal-specific revenue breakdowns are either available at BICS level 5 or BICS level 6.

Table 1: BICS sub-classifications for the Metals & Mining Group

By querying for BICS revenue using the appropriate BICS code override and equity fundamentals year, one is able to obtain annual revenue breakdowns (Table 2).

Table 2: Example BICS Revenue breakdown (Zimplats Holdings, 2018)

The US dollar denominated revenue is converted to an estimated production level using the average prices as detailed in Table 3.

Table 3: Commodity pricing source and annual averages (2011 – 2019)

Production Groups

A production group contains companies which have positive revenue/production in the same set of metals. From 2011 to 2019 we find approximately 45 such groups across all metals, far less than the number of theoretical combinations. Industrial metal intensities can now be estimated using least squares on the data in Figure 7 (main body of the publication) with GHG being the dependent variable and production levels as regressors.

Table 4: Industrial metal intensities

Unfortunately, applying the same least squares approach as above on precious metals results in a negative intensity for silver. A possible explanation of this can be obtained by returning to the distribution of company-level gold intensities as shown Table 5.

Table 5: Distribution of GHG intensities among pure gold producers

The mean, median and trimmed mean of the ‘middle 50’ are 25,772, 20,298 and 20,725, respectively. Note the median is less than the mean, which is consistent with the sample skewness of 1.72. The distribution of companies’ implied silver intensity under these assumptions for gold is shown in Table 6.

Table 6: Implied silver GHG intensities among Gold + Silver producers using an assumed gold intensity of 20,298

Of the relatively small sample size (20 observations), 4 data points have extremely large negative implied intensities associated with them. While it is difficult to be sure about the cause, these negative values are certainly problematic for a regression-based approach. Table 7 provides an indication about the dependency between estimates and the measurement metric selected.

Table 7: Silver intensity estimates – using trimmed mean versus median

Alternative Approaches

One of the first things we tried was regressing company emissions on company production. However, we consistently found a few negative estimates, regardless of attempts to winsorize data or use weighted least squares. We believe this seemingly intuitive approach does not work for a few reasons:

• Production levels are not enough to explain company-level variation; technological and geographical factors likely play a significant role in explaining emissions. In short, there are missing regressors for a company regression.

• It is unlikely that emissions are a linear function of production. For example, when a metal is produced as a byproduct of primary mining operations the marginal GHG emission is less than on a standalone basis. While we could have included interaction terms, hoping for a negative loading when metals are jointly produced, it is difficult to see what to do with these coefficients when determining a final per-commodity average.

As we saw with industrial metals, by adding all production and emissions within a production group, we eliminated the need for regression to explain company-level variations. It also seems to limit the (possibly) detrimental effects of cross-production in a linear model—since simple least squares then leads to estimates consistent with previous research. Our view is that for commodities investors, explaining aggregate emissions ought to be sufficient, and that a simple linear estimation is likewise preferred.

On the whole, silver seems to be among the most difficult metals for which to estimate GHG. This is likely due to the high variability in the mining and production processes. Silver is a common byproduct of other mining operations; in our data set, there are no stand-alone silver producers. As we saw above, even when restricting the universe to Gold + Silver producers, there is so much variability that any plausible average gold intensity would lead to at least some implausible silver intensities among companies (Table 6).