The Perils of PV Price Prediction

In the ongoing effort to deploy more sustainable energy technologies, there is no more gratifying sight than the now-familiar plot of PV module price declines that have occurred as the cumulative manufactured volume has grown, as illustrated in Figure 1. Plotted on a log-log scale, the values chosen for the average sales price tend to fall along a (somewhat) straight line. This downward trend is sometimes referred to as “Swanson’s law,” after Richard Swanson of SunPower, who noted the linearity in 2006.

The price decline evokes comparison to Moore’s law, which suggests that the behavior might be expected to continue into the future. Moore’s law famously, over the course of four decades, predicted the doubling of the number of transistors on an integrated circuit every one to two years. Though there is healthy debate regarding the extent to which the PV module price trend can continue, there is nevertheless a natural tendency to extend the line into the future as a way to predict future PV module prices.

Figure 1: Module Average Selling Price, 1980-2012

Source: IEA Technology Roadmap: Solar Photovoltaic Energy 2014

PV module price behavior is a composite of multiple underlying factors derived from both the supply and demand sides of the PV market. Continuation of the linear price trend requires that the net impact of these factors remains consistent. One supply-side factor is manufacturing experience, which often enables a reduction in the cost of manufacturing as volumes increase and manufacturing efficiencies are found and exploited. Another supply-side factor is the cost of materials: the polysilicon feedstock shortage caused increased prices during 2003-2008 that diverged from the linear trend. Manufacturing capacity expansion increased the supply and helped module prices to return to their previous downward trajectory after 2008 and into the present.

The effect of demand-side factors should also be considered. For most of the past three decades, PV has been relatively expensive; demand was therefore limited. While the PV market remains limited by demand, module ASP and manufactured volume are interdependent variables, making their behavior easier to predict. Plotting cumulative manufactured volume on the x-axis can create the impression that it is an independent variable, but the manufactured volume is itself dependent on the price. Just as increased manufacturing volumes may increase experience and thereby drive down module price, a reduction in price may just as readily increase demand and deliver an increase in the manufactured volume. As seen with the polysilicon shortage, any future supply shortages will likely disrupt this demand-limited behavior.

There are also several practical limitations in using Figure 1 to predict future price behavior. On a log-log scale, small differences in the slope of the line translate to significant differences in predicted future prices. Second, by collapsing the broad range that has existed in annual module prices down to a single average point, the validity of any resulting predictions becomes less certain.

To what extent, then, has past data from Figure 1 ever been predictive of future prices?

One way to assess this is to examine predictions retroactively, using various subsets of the data in Figure 1. For example, 20 years of data from 1980-1999 might be used to predict the ASP five or 10 years in the “future” (in this case, 2005 or 2010, respectively).

Figure 2(a) illustrates a series of such linear fits to the historic PV module average selling price (ASP), using data sets that increase in five-year increments. Linear fits are colored to match their respective data sets (colors go from cool to warm, indigo to red, as more recent data is added). Each increment in the data range causes a marked shift in the slope of the fit. When the linear fit is then used to predict the ASP five or 10 years beyond the last data point, the variation in these predictions is often pronounced. Looking ahead to 2020 and beyond, will a linear-fit prediction of ASP five years in the future be any better than a “guesstimate” based on recent ASP data?

To put the results in context, Figure 2(b) applies a similar approach to a data set drawn from the financial sector, one that is well known to confound efforts at prediction: the Dow Jones Industrial Average (DJIA). The percent differences of the predictions from actual values are compared in Figure 3. While the error in prediction for the DJIA is often higher than that for the ASP (particularly for 2010, in the wake of the global financial crisis), the predictive power of linear fits to past PV module ASP appears, at times, to be little better than that from the Dow Jones Industrial Average. As with the stock market, historical price behavior has tended to be informative only in hindsight.

As Yogi Berra said, “It’s tough to make predictions, especially about the future.”

Figure 2: Linear Fits Applied to Successively Larger Subsets of the PV module ASP (a, left) and the Dow Jones Industrial Average (b, right)

The indigo data points (1980-1984) were used to create the indigo fit lines, and so on. Data after 2009 is shown for reference; the last fits are for data through 2009 (red lines).