Maxim Kartamyshev, Senior Analyst, Norges Bank Investment Management
From the asset management perspective, the energy transition, i.e., substitution of fossil-based energy production with renewable energy sources, is a subject of significant interest. Energy transition policies lead, essentially, to structural and lasting changes in energy supply and demand. Estimating likely consequences of such changes for different sectors of the global industrialized economy is a non-trivial challenge.
At the macroeconomic level, the implications of an energy transition are investigated within the Shared Socioeconomic Pathways (SSP) research program[1]. The program aims at providing internally consistent integrated assessment scenarios, detailing hypothetical future developments in energy use, population, economic, climate change, and other relevant indicators. The online database of SSP scenarios serves as an excellent starting point for the exploration of investment risks and opportunities associated with different energy transition pathways.
It is apparent that the energy transition will affect all sectors of the global economy to varying degrees. Hence, the aftermath of changes in energy mix should be analysed with sufficient granularity, properly accounting for the economic interdependence of different sectors. This exercise requires a robust modelling framework.
To examine prospects for the sector-level energy transition analysis, we relied on the proprietary framework of the Economic Domains Model (EDM)[2]. In essence, EDM aims at modelling the world economy as an ensemble of interacting economic domains. The feasibility study considered only a dozen of such domains, corresponding to the largest sectors of the US economy. Using fundamentals of publicly listed companies to quantify domain-level economic activity, a collection of neural networks (NN) was trained to gauge impacts of energy transition on the target sectors.
Use of NNs as the primary modelling gear allowed us to consider a broad range of important transition variables. Of course, data augmentation techniques had to be employed to obtain large enough training datasets. Fortunately, the modelling context catered for a simple augmentation procedure, based on using clusters of companies within a domain to approximate domain-level economic indicators. With the models established, we constructed projections of sector-level fundamentals for energy transitions detailed by selected SSP scenarios.
Despite many approximations, the feasibility study produced compelling results. To identify the main transition variables influencing the projections, we employed Shapley Additive Explanations (SHAP) approach[3]. Enabling model-agnostic, unbiased feature attributions of ML forecasts, the approach also sparked interest in potential generic applications of interpretable machine learning for exploratory investigation of complex systems.
As universal approximators, NNs are well suited for quantitative descriptions of complex non-linear systems. Yet, success of the modelling is strongly dependent on the availability of sufficiently large datasets describing behaviour of the target system, alternatively on a realistic augmentation of available measurements. Within the data-driven approach, to improve generalization properties of NN models, one generally strives to obtain training datasets capturing essential system dynamics for a broad range of relevant regimes.
Equipped with generalizable NN model and adequate computational budget, it is tempting to apply SHAP to examine the inner workings of the target system. For example, one could compute explanations of NN outputs for a dynamic regime of interest, represented by a suitable grid in feature space. Observed relations between feature vectors and corresponding contributions to the forecasts could then be treated as phenomenological response functions, observed in a numerical experiment.
The extracted response functions, together with human expertise and intuition, are likely to assist in identifying hidden casual relationships, thus expanding available knowledge of the complex system in question. Serving as a numerical probe of dynamics captured by a black-box data-driven model, interpretable machine learning could hence provide a valuable supplement to explicit, analytical modelling toolbox.
Such numerical probes are particularly helpful for analysing the implications of extreme events. For example, it is straightforward to adapt an EDM methodology to estimate the impacts of lockdown policies on business activity in different domains. Yet, it is desirable to assess how well the resulting NN models generalize to the region of unprecedented changes in global economic activity and energy trends.
Recognizing that there is high uncertainty related to outcomes under the environment of ongoing lockdowns, we computed short-term estimates of sector-level developments for a large variety of hypothetical macro-economic futures[4]. Estimates of year-on-year (YoY) change rates in aggregated sales for selected sectors are presented on Figure 1.
Forecasted YoY change rates in sector level sales for hypothetical macroeconomic scenarios for year 2020. Each scenario is defined by corresponding YoY change rate in economic activity, energy consumption and electricity generation.
Oil & Gas Producers (US)
Automobiles & Parts (US)
Economic activity change rate [%]
Reviewing corresponding response functions (see Figure 2 for examples) revealed intuitively appealing relationships between changes in the macroeconomic indicators and forecasted levels of activity within different economic domains, encouraging further enhancements of the EDM framework.
Contributions to forecasted YoY change rates for year 2020 in sector level sales from selected feature groups, computed using SHAP methodology. Each scenario is defined by a corresponding YoY change rate in economic activity, energy consumption and electricity generation.
Oil & Gas Producers (US) - Contributions from macroeconomic features
Automobiles & Parts (US) - Contributions from macroeconomic features
Oil & Gas Producers (US) - Contributions from energy consumption features
Automobiles & Parts (US) - Contributions from energy consumption features
Clearly, to improve the realism of EDM forecasts, a much broader range of domains should be considered. Additional work is needed to properly account for interactions between domains of the interconnected global economy. It is also necessary - and possible - to increase the variety of relevant model features. For example, to accurately estimate the consequences of the inevitable changes in energy mix, we believe it will be helpful to utilize specialized measures such as energy return on investment[5].
Availability of a sufficiently realistic modelling framework could substantially improve prospects for the quantitative investigation of the ongoing energy transition. In addition to accurate and self-consistent analysis of potential consequences for hypothetical transition scenarios, the framework could possibly also be applied to identify transition pathways satisfying pre-defined optimality criteria.
Despite the numerous challenges associated with the process, we choose to remain cautiously optimistic on the prospects of enhanced EDM framework in attaining practical relevance for long-term investment management applications.
