Poster #101 - Forecasting learning rates of technologies for the design of effective industrial and innovation policy

• In Event: Saturday Poster Luncheon

Abstract

Understanding how the cost of nacent technologies decreases with deployment is fundamental for effective public policy design. This future cost reduction is estimated in the literature by the learning rate (LR), the percentage reduction in cost per unit for each doubling of cumulative deployment. The concept has now been applied to designing industrial strategies for industry localization and to tailoring innovation policy to support the development of new technologies. Our study aims to relate a technology's LR to its technology-inherent structural characteristics. The results can be used to ex-ante estimate the LR for a given technology. While existing literature already identifies structural characteristics as determinants of LRs, these studies are largely limited to energy technologies or constrained by small sample sizes. This work conducts the first large-N quantitative study of empirical LRs across diverse product sectors to answer the question: How are product-inherent characteristics correlated with LRs across sectors? 

Our database draws on an extensive literature review and includes 660 experience-rate data samples spanning a wide range of sectors, including industrial chemicals, consumer electronics, energy system technologies, and large infrastructure projects. In this study, the term “product” encompasses the full range of technologies, goods, and systems for which LRs have been empirically calculated. By prioritizing global data points with the longest observation durations, we refined the dataset to 201 representative LRs, each corresponding to a unique product. To systematically measure the role of structural characteristics on the LR, we classified the products along key characteristics theorized to affect cost reduction: unit size, design complexity, standardization, and natural resource-based industry (NRBI) dependency. While the production process is not inherent to the technology, we also test its mediating role on LR. 

Preliminary OLS regressions (n ≈ 190) show that when considered individually, unit size, design complexity, and standardization are each significantly correlated with LRs (p < 0.01) in the theoretically expected directions, whereas NRBI dependency is not. Logit regressions also show that for products that are large, complex, or customized, the probability of near-stagnant cost reductions with deployment (LR < 5%) ranges from 20–30%, which is much higher than 2-4% for small, simple, or standardized products. The lower LRs for large, complex, or highly customized products may be explained by their reliance on bespoke production processes, which require substantial inputs of tacit human knowledge that tend to remain localized, constraining the transfer of knowledge that drives cost reductions.

These results provide an empirical foundation for matching industrial strategy to the inherent structural characteristics of nascent technologies, and highlight an important policy tradeoff: where multiple technologies with different LRs are available for a given application, policymakers must weigh fast-learning technologies with difficult-to-localize benefits against slower-learning technologies whose production-embedded tacit knowledge can boost the industrial value-add of their domestic economy. Policymakers should therefore use structural characteristics not only to forecast cost trajectories but also to select the policy design that best matches each technology while delivering the desired trade-off between national and global policy goals.

Author

• Anurag Panda, ETH Zurich
  Presenting Author