Comparing expert elicitation and model-based probabilistic technology cost forecasts for the energy transition

The use of model-based approaches to estimate future technology costs has a long history. The various approaches can be understood in terms of three elements: a) the “underlying model” of technological change—that is, the functional form or relationship between dependent and independent variables; b) the way uncertainty is characterized and used to project the underlying model forward in time to generate forecasts; and c) the way forecasting method performance is evaluated using forecast error models.

To make probabilistic cost forecasts, we use two functional forms—Wright’s and Moore’s laws—plus two methods of projecting uncertainty forward, yielding a total of four model-based methods (see Methods for more details). The first uncertainty projection method we use is the first-difference stochastic model developed in refs. 27 and 29, which we call the Stochastic Shock method. This model assumes there is a stable but uncertain technological change trend, on top of which periodic stochastic shocks occur, impacting costs repeatedly over time. The model is calibrated using differences between sequential observations in each technology’s data series.

The second uncertainty projection method we use is a modification of the method proposed by ref. 25. The original method involves calculating the frequency with which learning rates (for Wright’s law) or progress rates (for Moore’s law) are observed in a given technology’s data series, using a combinatorial approach, and then using the resulting distributions to project uncertainty forward. However, using all start/end year combinations to calculate these distributions pools together observations that are not entirely comparable since ranges overlap and vary in length. To avoid potential biases related to double counting, we use only single-period learning and progress rate observations; we call this modified method the Stochastic Exponent method. We fit normal distributions to the observed progress exponent distributions and generate forecast distributions in a Monte Carlo method style. Each single forecast is made by sequentially picking progress exponents from the relevant normal distribution and projecting costs forward, year by year, until the time horizon is reached; then, these are aggregated to form the final forecast distribution.

Expert-based approaches, which include EEs, are a different class of forecasting methods (SI Appendix, Table S1). They can be used to offer insights into the future of technologies for which historical deployment and cost data are not yet available or commercialized (14, 68), are designed to always provide probabilistic estimates of future cost developments (69), and do not assume that previous trajectories will continue and do not preclude the identification of technology surprises or discontinuities.

Among the expert-based approaches, EEs have been the most widely used method to estimate probabilistic future technology costs in the energy sector. Because of this relative richness of data, we focus our analysis of expert-based methods on EEs. EEs are structured surveys of experts who are asked to provide probabilistic estimates of future costs by using the best available information to them at the time of the elicitation (16, 32), which is likely to include different types of observed data. It is important to note that in some cases, this same observed data may serve as input to some of the model-based methods.

The EE method was pioneered in the 1960s and 1970s to support decision-making in the presence of extreme or unlikely events (70). EEs have often been applied to forecast technology cost and performance to support decision-making under uncertainty (16). In the past decade, an increasing number of expert elicitation studies were conducted on energy technologies (14) including (among others) nuclear power (71), wind energy (32), solar energy (72), water electrolysis (73), and energy storage (74).

EEs are of particular interest when data are sparse or missing or the technologies are emerging. However, they are time consuming and may differ in terms of the elicitation protocol and methods for administering the survey (Dataset S1). Research has suggested that in-person interviews have the advantage of allowing experts and researchers to have in-depth discussions and ask clarification questions but may be subject to small sample sizes and cognitive biases (72). Other research has shown that although online instruments are often pilot tested and can enable a broader representation of experts, which can reduce biases (33), challenges can remain since questions cannot be tailored in real time to adapt to respondent preferences and understanding. Some effort has also been devoted to understanding quantitatively the EE design factors that lead to systematic differences in uncertainty estimates from energy technology expert elicitations. For example, ref. 33 explores the role of elicitation mode, expert characteristics, and research and development (R&D) scenario variables in determining the uncertainty range surrounding energy technology cost forecasts using the largest set of EE studies available as of 2015. In the same vein, other work has investigated the extent to which elicitations focused on aggregate- versus component-level costs are associated with differences in elicited cost forecasts (75, 76). Another important consideration related to EE design is that, given that experts are busy and have limited time available, they are necessarily asked to provide estimates about a limited number of points in time. This means that using EE forecasts for specific policy or investment decisions and comparisons (such as the ones in this paper) often involves interpolating or extrapolating elicitation data over time.

Beyond questions about elicitation design, using EE outputs often requires making difficult decisions about whether or not to aggregate expert elicitations and, if so, how. While offering insights that are more easily interpreted by policymakers and other analysts, aggregating across experts presents limitations and drawbacks that must be considered when it comes to the communication of results as discussed in refs. 76 and 14. The aggregation, interpolation, and extrapolation methods used in this work are discussed in Methods.

Despite the significant and growing research on model-based and expert elicitation forecasts in energy technologies and other technological areas, we are not aware of any research collecting and systematically comparing the performance of those types of approaches to observed values or each other in the future.

The first step and contribution of our analysis is a large collection of data on energy technology costs, which we make available as detailed in SI Appendix, section 3. These include 25 sets of data from EEs on energy technologies conducted between 2007 and 2016 and covering a range of geographies and 25 sets of observed energy technology data including cost, deployment, and time. Dataset S1 lists all the EE data sources by technology, including information about how the data points were described in the original sources and the cost metrics available for each technology. Dataset S2 lists all the references and links for both the EE and observed data. Note that due to the need for both elicitation and sufficient prior observed data on the same technology, forecast comparison analyses were only possible for much smaller subsets of technologies than those in the complete data set.

The second step of our analysis is to generate probability distribution functions of estimated costs from model-based and expert-based methods as detailed in Methods. This allows us to compare the probabilistic 2019 cost forecasts from both methods to observed average 2019 costs as well as to compare the 2030 probabilistic cost forecasts generated using the different methods to each other. Here, we highlight a few key aspects of the forecast generation process.

To build the expert elicitation forecasts for the comparison we a) fit one of three types of continuous probability distributions to the discontinuous probabilistic data points provided by each expert, b) aggregate the resulting continuous individual expert distributions for each technology assigning equal weights to each expert using the method in ref. 26, and c) when the time points provided by the experts did not coincide with our 2019 and 2030 forecasting horizons, we use an exponential fit between the two time horizons provided by the experts to either interpolate or extrapolate as required (SI Appendix, section 3).

To build the model-based forecasts, we generate probabilistic estimates from Wright’s law using the Stochastic Shock (W1) and the Stochastic Exponent (W2) methods and from Moore’s law using the Stochastic Shock (M1) and the Stochastic Exponent (M2) methods (Table 1). Three important features of the model-based forecasting process should be mentioned.

First, to apply Wright’s law model, one needs to make assumptions about the future level of deployment. In this work, we assume recent historical growth trends continue into the future and persist for the duration of the forecasts: we extrapolate deployment using the compound average annual growth rate (CAAGR) observed over the 10-y period prior to each forecast. The W1 and W2 forecasts are therefore conditional upon these specified levels of deployment. We refer to this as a “continuation of past trends deployment scenario,” which is tantamount to saying that deployment, R&D funding, and other variables continue the historical trajectory of the previous 10 y. We depart from this general rule in the case of two technologies because the calculated average rate of deployment over the 10-y time period may be too high compared to even the most ambitious deployment scenarios for those two technologies (see Methods).

Second, in the computation of the 2019 model-based forecasts, we only use observed data up to the year of the elicitation with which the forecasts are to be compared. In other words, if the elicitation for a given technology was conducted in (let us say) 2010, we produce the model-based forecast of 2019 costs using observed data up to 2010 only, even if observed data are available beyond 2010. This allows us to compare the EEs with model-based estimates that do not enjoy the benefit of additional years of data that were not available to the experts when they made their forecasts.

And third, although in many cases we found that relevant observed data were not available to generate model-based forecasts, in other cases we had to make choices about which data to include. The details on the data selection for nuclear power, offshore wind, and two electrolysis technologies can be found in SI Appendix, section 3.

Fig. 1 presents the systematic comparison of expert-based and model-based probabilistic forecasts with observed 2019 costs. The figure shows average 2019 observed costs and the probabilistic 2019 EE, W1, M1, W2, and M2 forecasts for the six energy technologies for which both elicitation and sufficient prior observed data are available. The horizontal axis shows the cost and the cost units, and the different colors on the vertical axis denote the forecast method. The whiskers represent the fifth to 95th percentile range, the short sides of the rectangular box the 25th to 75th percentile range, and the line dividing the rectangular box the 50th percentile (the median). The corresponding average observed 2019 cost is shown with a dashed gray vertical line in each of the six plots.

From left to right and from top to bottom, the technology forecasts are listed starting with the oldest forecast. The oldest forecasts are for nuclear power, which use an expert elicitation from 2009 and, correspondingly, observed data up to 2009 to produce the model-based forecasts. The most recent forecasts are for alkaline electrolysis cells (AEC) and proton exchange membrane (PEM) electrolysis cells, which are based on an expert elicitation from, and observed data up to, 2016.

Fig. 2 shows the log of the ratio of the 50th percentile of the cost forecast and the observed cost value. This is a dimensionless metric for characterizing the performance of the different methods, which allows us to systematically compare the size of the uncertainty ranges and the distance between the forecast medians and the observed average values. The line encompasses the log of the ratio of the 95th and fifth forecast percentiles and the observed value.

There are three main takeaways from the performance comparison shown in Figs. 1 and 2.

First, we find that, for this set of energy technologies, model-based forecasts outperform EE-based forecasts in terms of their ability to capture the 2019 observed value within the forecasted 2019 uncertainty range. Concretely, almost all model-based approaches produced fifth to 95th percentile ranges for 2019 that included the observed values for all six technologies. The only exception was the M2 method on solar PV, which failed to include the observed value, although only by very little (Fig. 2). Notably, the Stochastic Shock method (W1, M1) produced 25th to 75th percentile ranges that captured the average 2019 value in all technologies but PV. In contrast, the fifth to 95th percentile range of the EE 2019 forecasts only contained the observed cost value for the two wind technologies (i.e., onshore and offshore wind). Model-based methods also outperformed elicitations in that they produced forecast medians that were closer to the observed 2019 values. Specifically, model-based medians were closer to the observed costs than the EE medians for all technologies except offshore wind and the W2 model-based forecast for PEM electrolysis.

Second, in almost all cases, both the EE and the model-based forecast medians are higher than the observed 2019 costs—they underestimated technological progress over this forecasting period. The notable exception is nuclear power, in which the finding is reversed—all forecast medians are lower than the observed cost. The W2 forecast medians for the two electrolysis technologies are outliers; reference SI Appendix, section 4 for a detailed discussion of the uncertainty ranges generated by W2 and M2 in this case.

Third, for four out of the six technologies, the fifth to 95th percentile range of the 2019 forecast was larger for the model-based methods than for EEs (Figs. 1 and 2). This was not the case for solar PV and onshore wind. Note that the fifth to 95th percentile ranges for EE and model-based forecasts for onshore wind were very close. The fact that model-based forecasts tend to generate fifth to 95th percentile uncertainty ranges that are generally larger or equal to those of EEs may partly explain the first finding discussed above. It is important to point out that the EEs all included steps in the elicitation protocol aimed at reducing expert overconfidence.

Determining why all methods underestimated technological progress in all technologies except nuclear when generating 2019 forecast medians is hampered by the fact that our analysis was possible only for a small number of technologies. However, we can hypothesize a few possibilities. These include unforeseen changes in the industry structure and/or policy and/or an underestimation of the level of deployment. The five energy technologies for which forecast medians were higher than observed 2019 costs are likely correlated, that is, subject to a common set of underlying drivers related to decarbonization policies, investor perceptions and preferences, firm expectations, and societal pressures. It is therefore possible that the forecasting period used for our systematic comparison was one in which a range of underlying drivers of innovation changed to result in faster innovation across those five technologies. Along similar lines, for elicitations, under- or overestimation of technological change can occur if experts cannot foresee increases or decreases in policy support, deployment, or regulation for this particular set of technologies. As previous research suggests, it is also not surprising to see nuclear evolve differently (52, 77, 78) from other technologies. In other words, this faster pace of innovation in most of the energy technologies covered compared to the forecasts is likely the result of structural change across the energy sector due to widespread policies and social and market forces.

Importantly, our results do not necessarily imply that EE forecasts will be less accurate than model-based forecasts for all technologies. Indeed, SI Appendix, section 5 and Fig. S3) shows three additional EE technology cost forecasts (bioelectricity, lithium-ion batteries, and biodiesel) whose fifth to 95th percentile ranges included the 2019 observed values. These are not shown in the main body of the paper because no observed data were available to generate comparable model-based forecasts. It is of course possible that, had model-based forecasts been available, they would have been more accurate than these EE forecasts, or vice versa. In SI Appendix, section 6, we include an additional robustness check with different observed data inputs for the model-based forecasts for offshore wind—not including experience from onshore wind. The results are consistent with those shown in Figs. 1 and 2.

This comparison of different methods’ performance represents only a first step, yet an important one. As more data becomes available, future research could further test whether model-based approaches outperform EEs in other technologies building on the approach presented here. We note that the ability to compare the performance of forecasts from different methods is directly determined by data collection efforts, pointing to the need for continued data collection to enable comparisons for a larger group of technologies. Two other important subjects of future research are, first, how to develop model-based forecasts for technologies that may be correlated due to structural changes in the underlying markets, and second, how to further reduce overconfidence in EEs.

We also compare probabilistic EE and model-based 2030 cost forecasts with each other (as 2030 costs are not yet known) for the 10 energy technologies for which the necessary data are available. Fig. 3 shows the median and the fifth to 95th percentile ranges of the W1, M1, and EE forecasts to the year 2030 for onshore wind, offshore wind, crystalline silicon PV modules, thin-film PV modules, concentrating solar power (CSP), all PV modules, bioelectricity, nuclear power, AEC, and PEM electrolysis cells. 2030 forecasts using W2 and M2 forecasts are not presented here to increase readability, but they are shown in SI Appendix, section 7 and Fig. S5.

To generate the model-based forecasts for 2030, we rely on all the data available as of late 2020—that is, we do not shorten the time series to match the year of the expert elicitation as we did in Figs. 1 and 2. The objective of the comparison to 2030 is to summarize what is known today about the possible range of 2030 costs and to identify differences between EE and model-based forecasts going forward. Fig. 3 includes two different estimates for crystalline silicon (c-Si) PV costs, one relying on the full set of observed data for c-Si PV (Fig. 3C) and one relying on observed data since 2006 only (Fig. 3D, Left). We do this because the data available for thin-film PV covers only the period 2006 to 2019, so the most meaningful comparison with c-Si may be one that uses data beginning in 2006 also. This illustrates the importance and nuance of data choices underpinning the model-based forecasts. And as previously noted, SI Appendix, section 3 contains details regarding the data used as input for model-based methods in the case of offshore wind. We also conducted a sensitivity analysis with two additional deployment scenarios: one in which we use a deployment scenario consistent with the International Energy Agency’s (IEA) Stated Policies Scenario and one that is consistent with the IEA’s Sustainable Development Scenario (reference SI Appendix, section 8 for more information). Fig. 4 compares 2030 cost estimates of those subtechnologies that are more mature and established (i.e., “dominant”) with those that are typically emerging, or with a much lower market share (i.e., “novel”).

Three main insights emerge.

First, the uncertainty ranges (i.e., the 5th to 95th percentile range) for the 2030 forecasts generated using W1 and M1 are generally larger than those for the EE forecasts. This is similar to what was observed in Figs. 1 and 2 for the 2019 forecasts. The results for W2 and M2 are largely consistent (SI Appendix, section 8 and Fig. S5 and related discussion). While we cannot yet determine the accuracy of these forecasts when compared to observed 2030 costs, the smaller 2030 EE forecast ranges may be less likely to include them. Given the focus in the expert elicitation literature on addressing overconfidence, the fact that EE forecasts generally have smaller uncertainty ranges compared to those from model-based methods again suggests that additional research to reduce overconfidence would be useful.

Second, for nine out of the 10 technologies, the model-based 2030 cost forecasts have lower medians than the EE forecasts; nuclear power is the one exception in which EE forecasts are lower than model-based forecasts. For all technologies except bioelectricity, the EE forecast medians in 2030 are lower than the observed cost of the technology in the year when the EE was carried out (listed in the gray bands in Fig. 3), reflecting a general expectation from experts of cost reductions, although with substantial differences across technologies. The W1 and M1 forecast medians in 2030 are lower than the 2019 average costs for all technologies except nuclear power (and, in the case of W1, AEC also, though the difference is small). Reference SI Appendix, section 8 for more discussion on the uncertainty ranges generated using W2 and M2. Importantly, the W1 (and W2) findings presented here are robust to the different deployment assumptions used in the sensitivity analysis as discussed in SI Appendix, section 8.

Third, as illustrated in Fig. 4, “novel” subtechnologies generally have higher 2030 EE forecast medians and are characterized by larger uncertainty ranges than the EE forecasts for the corresponding dominant technologies. The comparison across all subtechnologies using all five forecasting methods was possible only for wind power, for which the comparatively novel subtechnology (offshore) has achieved significant diffusion already. This fact points to the particular value of EEs as a source of information to get an initial understanding of the future costs of technologies for which cost data are not available. Only as time progresses will we get a better sense of whether or not these comparatively novel technologies will “catch up” with their currently dominant technology counterparts by 2030 in terms of the costs. See more discussion in SI Appendix, section 9.

The results discussed suggest that EE 2030 forecasts should not be used for those technologies that have experienced significant cost reductions between the time of the expert elicitation and 2019. For several of the technologies covered in Fig. 3 (specifically Fig. 3 A, C, D, and F), the 2019 observed costs are already lower than the 2030 EE median forecasts. In these cases, model-based forecasts, which reflect information from the last 6 to 10 y, are preferable.

One possible explanation for the fact that most model-based forecasts were lower than EE forecasts emerges from the literature on the role of technology modularity or complexity (52, 78, 79) as determinants of technology innovation trajectories. Large-scale nuclear power plants and (to some extent) bioelectricity and CSP plants would fall into the category of less modular technologies compared to solar panels and batteries, for instance. Model-based 2030 cost forecast medians are lower than those from elicitations for more “modular” technologies and higher for the least modular technology (nuclear). They are closer to each other for the two technologies that are arguably in a midrange of modularity (CSP and bioelectricity). However, statistically testing this hypothesis is not within the scope of this paper. The fact that experts expect faster cost reductions for less modular technologies compared to model-based methods may also be due to the following: additional information they may possess about scientific breakthroughs, the industry, and/or policies; unawareness about the correlations between modularity and technological change; and/or bias and overconfidence. This indicates the need for future work to further test this finding using different metrics to account for modularity and/or complexity (e.g., ref. 79) as more data becomes available.

To produce probabilistic technology cost forecasts, we collected two types of data. First, we collected data on the evolution of cost and performance of energy technologies over time along with installed capacity or cumulative production. These are required to generate the model-based forecasts of future costs. We obtained this data from a range of databases (e.g., International Renewable Energy Agency [IRENA]), research articles [e.g., Nagy et al. (28)], and research organization reports [e.g., Fraunhofer (82) and Lawrence Berkeley National Laboratory (83)]. Second, we collected forecasts of future costs and performance (typically around 2030) using EEs from academic publications e.g., Verdolini et al. (14)]. Dataset S2 includes an overview of the technologies for which observed data were available (a total of 32 observed datasets in energy technologies), the time period covered, and the source.

We first describe the two underlying technological change models used in this paper, and then we describe the two uncertainty characterization methods and then finally, the expert elicitation forecast generation method.

Wright’s law postulates that cost decreases at a rate that depends on the cumulative production as described by the following:

in which $y_t$ is the technology cost in year t, A is a constant, X_t is cumulative production, and w is the Wright exponent (or learning exponent). This exponent is then used to calculate the “learning rate,” defined as r = 1 − 2^-w. This rate is interpreted as the percentage reduction in costs associated with each doubling of cumulative production.

Moore’s law describes the exponential decrease in cost y_t of a technology as a function of time according to the following:

in which B is a constant, t is the year, and $\mu$ is the Moore exponent (or progress rate).

This method uses the difference equation form of Wright’s law (where t stands for time in years):

and of Moore’s law:

To implement the Stochastic Exponent method, we first calculate all interannual exponents observed in the data for a particular technology (Wright exponents for W2 and Moore exponents for M2). For example, if we have 11 y of observed data for a technology, we can generate 10 exponents for both Wright’s and Moore’s law. Then, for each law, we fit a normal distribution to the sample of exponents obtained to create a Wright exponent distribution and Moore exponent distribution. These calculated exponent distributions are then used to generate cost forecast distributions for the technology by simulating large numbers of cost sample paths (we used 10,000 per forecast). Each sample path is generated by sequentially picking exponents from the corresponding exponent distribution for each year of the forecast and applying either the Wright’s law or Moore’s law difference equation, as shown above, until the forecast horizon is reached. The sample paths are then aggregated to approximate a single probabilistic forecast in the target year.

For the application of W2, it is necessary to assume a deployment scenario. In most cases, we used the CAAGR of cumulative experience (either installed capacity or electricity generation) observed over the 10 y before the forecast year to specify a future deployment scenario. There are two exceptions though: the PV and wind 2030 forecasts; in these cases, we used the CAAGR observed over the most recent 5 y instead (SI Appendix, section 8). This method implicitly assumes that deployment, R&D funding, and other variables continue on their recent historical trajectories for the entire duration of the forecasting period. We conducted a sensitivity analysis for the Wright’s law forecasts using two additional deployment scenarios as detailed in SI Appendix, section 8.

To generate the model-based forecasts for the 2019 comparisons, we only use observed data up to the year of the elicitation with which the model-based forecast is to be compared, but for the 2030 forecasts, we use all available data.

Importantly, the model-based forecasts using both the Stochastic Shock and Stochastic Exponent methods can be produced only in those cases for which we have at least ∼6 y of sequential annual technology data. This is because these methods infer cost trends from interannual cost differences, and at least around five samples are required for model calibration.

The Stochastic Shock method was developed and statistically tested by Farmer and Lafond (27) and by Lafond et al. (29) for Moore’s and Wright’s laws, respectively. The model represents the idea that in each year, technological progress occurs according to a stable underlying trend (either Moore’s or Wright’s law), and, in addition to this cost development, there is an exogenous shock that also impacts the cost. The magnitude of both the underlying trend and the stochastic shocks are specific to each technology and are determined by calibration using observed data. The periodic random shocks accumulate over time, giving rise to a probability distribution of forecast costs that grows wider over time (in log space, though not necessarily nonlog space, depending on the magnitude of the progress trend). In the simplest version of the model, the periodic shocks are independent and identically distributed (I.I.D.), but we use an augmented version in which they are correlated from one period to the next.

For Moore’s law, the model is as follows:

in which μ is the Moore exponent, ${\upsilon }_t\sim N\left(O,{\sigma }^2\right)$ are I.I.D. noise terms, and $\theta$ is an autocorrelation parameter. Following (27), we set $\theta$ = 0.63.

For Wright’s law, the corresponding model is as follows:

in which w is the Wright exponent, $u_t$ ∼N (0,σ_u²) are I.I.D. noise terms, X is the future deployment scenario (which is the same as that used for W2 as described in the previous section), and $\rho$ is the autocorrelation parameter. Following ref. 29, we set $\rho$ = 0.19. For a given technology, the parameters are estimated by using the observed data to perform an ordinary least squares regression through the origin. As such, the calibration relies on differences between sequential data points, which limits the data sources available for analysis (since many sources provide nonsequential data).

A hindcasting procedure was implemented in refs. 27 and 29 to statistically test the ability of these models to forecast observed progress trends. For this, a dataset of more than 50 technologies was used, spanning a variety of forecasting periods. For each model, many subsamples of data were used to make many forecasts for all technologies, and the resulting forecast errors were pooled to form an aggregate forecast error distribution. This was then compared to the forecast error distribution expected to arise from the model. In these two papers, the empirical and theoretical error distributions were a close match, indicating that the models did a good job at forecasting out-of-sample progress trends. Global values of the autocorrelation parameters ($\theta$ and $\rho$) for all technologies were estimated using the pooled sets of forecast errors.

Finally, refs. 27 and 29 derived analytical expressions for the forecast error distributions implied by each model in terms of all the estimated parameters. In this paper, we use the observed data for each technology to estimate all relevant parameters and then calculate the required percentiles of the probabilistic forecasts using these analytical expressions. The resulting forecasts thus take account of the underlying progress trends and observed volatility in different technologies’ historical records as well as the future deployment scenario (for Wright’s law) or time horizon (for Moore’s law).

It is important to note, first, that there are not large numbers of EEs in energy technologies [the first one we are aware of was published in 2008 (84)]. Second, many EEs only present cost estimates for one point in time, making it impossible to observe how cost distributions are expected to change over time. However, this is required in order to infer a distribution of costs at some intermediate year. Thus, we can only generate probabilistic 2019 cost forecasts for the subset of the expert elicitation studies that included information on technology costs in at least 2 y (Dataset S1).

In order to obtain estimated expert elicitation forecasts in 2019, we interpolate costs between 2010 and 2030 (nuclear and PV) or extrapolate costs between 2020 and 2030 (wind and water electrolysis) using an exponential functional form. This step is necessary for the comparison with model-based forecasts and realized costs in 2019 (reference SI Appendix, section 3 for further details).