Dynamic Evaluation of Policy Feasibility, Feedbacks and the Ambitions of COALitions

10 While the Paris Agreement instituted bottom-up coordination into international climate negotiations, 11 state-of-the-art integrated assessment models (IAMs) implement policies from the top-down, 12 distributing burdens subjectively or normatively. Here, we introduce the first evidence-based approach 13 for emulating real-world policymaking, Dynamic Policy Evaluation (DPE). Just as IAMs rely on empirical 14 relationships to prospectively quantify myriad techno-economic variables and simulate investment 15 activity, DPE endogenises national policy adoption based on observed causations between IAM 16 variables and political decisions. We demonstrate DPE on the Powering Past Coal Alliance (PPCA) via 17 iterative feedback loops between the IAM REMIND and a policy feasibility model, deriving probabilistic 18 scenarios with multi-stage accession. Our scenarios estimate baseline ambition toward “consigning 19 coal to history,” the 1.5°C-consistent entry point prioritised by COP26, exposing the potential loophole 20 of non-electric coal demand and other carbon leakage risks. We then assess path-dependencies of 21 PPCA expansion to Covid-19 recovery actions, illustrating DPE’s utility for exploring policy interactions.


25
Under the Paris Agreement, 175 nations agreed to common-but-differentiated responsibilities toward 26 limiting global warming to 1.5-2 o C above pre-industrial levels 1 . While cost-effectiveness analyses (CEA) 27 by integrated assessment models (IAMs) derive techno-economically and geophysically feasible 28 pathways to achieve the climate targets 2,3 , the political feasibility of these scenarios is under scrutiny 4-29 7 . Socio-political barriers are well-acknowledged, typically analysed through exogenously-determined 30 'second-best' scenarios, such as delayed action 8 , regionally-differentiated ambition 9 , or technological 31 skepticism 10 . However, these still presume global policy coordination, which appears infeasible in a 32 bottom-up international regime without credible enforcement mechanisms 11,12 . 33 Whereas CEA explores the political ambition needed to achieve stated goals, stated policy evaluation 34 (SPE) illustrates the consequences of maintaining current ambition levels, e.g. already-implemented 35 national policies (NPi) or nationally-determined contributions (NDCs) to Paris. SPE scenarios are often 36 used as reference baselines for CEA and policy evaluation analyses (PEA), which assess subsequent 37 mitigation options for their potential contribution to specified targets (Table 1). Conspicuously, for all 38 the endogenous techno-economic dynamics represented in IAMs 13 , SPE and PEA rely on exogenous 39 assumptions to prescribe policies top-down across disparate societies. To portray realistic expectations 40 for baseline ambition and subsequent policy options, models should instead emulate the bottom-up 41 nature of climate politics 14,15 . Two methodological innovations are necessary to achieve this: (i) to 42 objectively and dynamically quantify policy feasibility 6 and diffusivity 16 , and (ii) to harness bidirectional 43 feedbacks between national policy adoption and the global energy economy 7 . 44 Here, we introduce dynamic policy evaluation (DPE), a novel IAM approach (Table 1) which fulfills both  45 requirements to endogenise bottom-up policy coordination. Given that IAMs derive long-term energy 46 system investment patterns consistent with empirical data and anticipated socioeconomic trends, it 47 follows that observed policy developments can be coherently extrapolated in parallel. Recent empirical 48 research has begun to codify causal links between national techno-economic contexts and real-world 49 political decisions [17][18][19] , and vice-versa 20 . DPE merges 7 this knowledge with SPE. To wit, SPE captures the 50 global energy system impacts of an emerging policy initiative in the variables computed, which can be 51 input to empirical models that then systematically define policy stringencies across model regions and 52 periods for a subsequent scenario (Methods; Figure M2). This iterative feedback loop mimics the co-53 evolution of energy economics and energy politics; each government's behavior can be influenced by 54 the actions of any other nation (s)  As global systems and national politics co-evolve, where will coal phase-out policies become politically feasible, and how much coal can be expected to phase-out by 2050?
Concurrent endogenous assessment of a policy's techno-economic feasibility via IAM and political feasibility via empirical analysis of IAM scenario data. This interdisciplinary coupling captures reciprocal feedbacks between policy adoption and the energy system, improving realism of future policy uptake and thus emissions. of joining the PPCA. Specifically, the study analysed a pool of 2,036 regression models, permuting 93 eleven independent variables seeking to explain PPCA membership, and established that high per-94 capita GDP and low reliance on coal for electricity supply (coal-power-share) have particularly strong 95 explanatory power (Figure 2a) 18 . In a first attempt to quantify future policy feasibility, we use the IAM 96 REMIND 13 to provide scenario data to the DFS via the novel COALogit model, which employs spatial 97 downscaling routines and probabilistic thresholds, or 'socio-political tipping points' 37-39 , within the 98 PPCA-DFS to iteratively define country-level, evidence-based scenarios of PPCA growth for REMIND 99 analysis ( To address these questions, we model 18 scenarios investigating three dimensions: coalition 128 expansion, policy ambition, and Covid-19 recovery ( Table 3). The REMIND-COALogit model-coupling 129 framework mimics the PPCA's staged accession through an iterative cascade ( Figure M4) which 130 dynamically fragments policy stringency across model regions. We first analyse the energy system 131 impacts of our 'median-estimate' probable-neutral scenarios alongside the analogous probable-brown 132 scenarios, selected for the divergence in China's behavior ( Figure 2c+d): 133 Thereafter, we analyse sensitivities across each dimension using efficacy indices for coal phase-out and 138 climate mitigation which compare scenarios on unit scales, where 0 represents reference (NPi) coal 139 consumption or CO2 emissions and 1 corresponds to 1. Demand-exit Unabated coal consumption phase-out along same timeline, except metallurgical coal is permitted a ten-year delay (2040 and 2060 deadlines) to reflect steel decarbonisation inertia and China's 2060 carbon neutrality pledge 40 .

Neutral (N)
Covid-19 recovery plans re-confirm national historical tendencies in terms of project completion rates and mean plant lifespans in the coal power sector until 2025.

Green (G)
Completion rates fall 50% and all shelved pre-construction projects cancelled, but plant lifespans unaffected.

Brown (B)
Project cancellation rates decline 50%, and plants operate 5 years longer than historical national average.

Reference Scenario
NPi(-covid) Currently-implemented national policies, a revealed-ambition scenario serving as our baseline. We model four variations: NPi-N, NPi-B, and NPi-G, which correspond to each Covid recovery scenario, and NPi-default, without Covid constraints.

NDC(-covid)
Stated-ambition scenario assuming full compliance with the first-round 'nationally-determined contributions' to the Paris Agreement. We model three Covid-dependent variations (NDC-N, NDC-B, NDC-G).

WB-2C
'Well-below 2 o C', a scenario with >67% likelihood of limiting global mean temperature rise to <2 o C above pre-industrial levels throughout the century. Without Covid constraints.
Hi-1.5C 'Higher 1.5 o C', a scenario with >50% chance of achieving the 1.5 o C target in 2100 with a moderate allowance of temporary mid-century temperature overshoot. No Covid constraints.

1.5C
Scenario with >67% probability of achieving 1.5 o C and a 50% chance of temporary overshoot by <0.1 o C. Along with NPidefault, used to define efficacy indices (Figure 4). No Covid constraints. to 1850GW globally (Appendix I), corresponding to a 0.8EJ/yr reduction in coal-fired power generation. 150 The resulting trends in national coal-power-shares and the general upward movement of per-capita 151 GDP along the 'Middle-of-the-Road' SSP2 41 development trajectory lead 45 of 48 OECD+EU nations ii to 152 exceed a 50% accession probability by 2025 (Figure 2b). COALogit assigns these nations to the 2p-N 153 coalition, and the power-2p-N REMIND scenario applies the power-exit policy to them in 2030. 154 Non-OECD+EU 2p-N Accession by 2045 155 Using results from these intermediate REMIND scenarios (Table M2)

203
At the global level, the power-2p-N policy-coalition scenario reduces coal use by 450EJ compared to 204 NPi-N. Indexed to NPi-default, this achieves just 1.2% of the cost-efficient coal phase-out derived in 205 the 1.5 o C scenario. Thus, the median-estimate power-exit scores just .01 on the coal-exit efficacy index 206 (Figure 4). The climate mitigation efficacy is even lower, scoring .01 (saving 6GtCO2). Still, these are 207 considerably better outcomes than power-2p-B, which underperform NPi-default on both indices (-.02  208 and -.01, respectively), implying that a global brown recovery from the Covid-19 recession may 209 outweigh the PPCA's long-term coal and emissions reduction prospects. In any event, the verbatim 210 power-exit contributes negligibly toward Paris-consistent abatement, assuming weak strengthening of 211 global carbon pricing and non-electric sector regulations. 212

214
For the demand-exit, COALogit returns a 2p-neutral coalition scenario identical to power-2p-N. These 215 182 members comprise 81% of global coal demand in 2020, 25% of which was from OECD frontrunners. 216 The demand-2p-brown coalition contains just one fewer member than power-2p-B (Serbia), totaling 217 179 nations which comprise 32% of 2020 coal demand. OECD members represent a 60% share. 218

229
The response of free-riding nations in demand-2p-N and demand-2p-B follow similar temporal profiles, 230 albeit with high variance in magnitudes (Figure 3c+d). Free-riders also increase industry electrification 231 and gasification ( Figure SF4e), but fuel it with coal ( Figure 3c). A knock-on coal-for-oil swap in extra-232 coalition transport liquids is evident following the OECD phase-outmuch stronger when China 233 freerides in the brown recoverybut inverts after non-OECD adoption. Coal drives the entirety of 234 extra-coalition carbon leakage in demand-2p-B (7% rate), which is just 24% of global carbon leakage 235 (30% rate). In demand-2p-N, free-rider leakage rates are slightly net-negative (-1% coal, -0.4% carbon), 236 so intra-coalition emissions are the sole driver of the 18% global carbon leakage rate. 237

238
Globally, the demand-2p-N scenario results in a coal phase-out of 10,300EJ from 2020-2100 compared 239 to NPi-N. Isolated from other policies, this 50%-probable Alliance leads to a cumulative 3040GtCO2 240 globally, saving 790Gt compared to NPi-N. Hence, moderate growth of a demand-exit coalition leads 241 to efficacy indices of .52 for coal phase-out and .22 for mitigation. China's abstention is highly 242 detrimental, as demand-2p-B scores .29 and .12, respectively. In both cases, the adverse effect of O&G 243 leakage is evidenced by the ~250% spread between coal and emissions abatement efficacies. The 95%-probable 1p and 5%-probable 3p coalition scenarios embody the considerable uncertainty 256 inherent to estimating future political decisions. For the demand-neutral case, the 1p-3p range of coal 257 phase-out efficacy is .05-.85, and .02-.37 for emissions mitigation (Figure 4). Power-neutral scenarios 258 have an uncertainty range of -.01-.02 for coal and -.01-.01 for emissions. Therefore, while the 259 demand-exit is highly sensitive to coalition size, the power-exit is robustly inconsequential. 260

Carbon Leakage 261
Carbon leakage primarily emerges through coal markets in power-exit scenarios and through inter-fuel 262 substitutions in demand-exit simulations. We find power-1p scenarios to be extraordinary cases which 263 exhibit >100% leakage rates (237% in power-1p-N). Figure SF4a suggests that the power-exit retards 264 electro-mobility learning, leading to lock-ins of inefficient CtL and oil. This (small-magnitude) feedback 265 is robust to coalition size but becomes overshadowed by other responses, resulting in a 56% carbon 266 leakage rate in power-3p-N. 267 Comparatively, the demand-exit tempers leakage: 72% in demand-1p-N and 17% in demand-3p-N. 268 Irrespective of policy choice, we find that global carbon leakage rates decrease as the coalition grows, 269 and intra-coalition leakage dwarfs extra-coalition leakage with sufficiently large policy uptake (all 2p 270 and 3p). These findings are all robust across Covid recovery scenarios. 271

Low-Carbon Substitution 272
The impact of the power-exit on VRE ranges from -3EJ in 1p-N to 348EJ in 3p-N. The decline in 1p VRE 273 penetration is another consequence of the negative electro-mobility feedback. Bioenergy and other 274 low-carbon energy (Bio&LCE) deployment experiences marginal upticks of 2-55EJ (1p-3p). Under a 275 demand-exit-neutral regime, these second-order effects range from 112-2070EJ for VRE and 63-1320EJ 276 for Bio&LCE. 277

278
We demonstrate that the demand-exit policy is 38x as effective at phasing out coal and 27x as potent 279 at CO2 abatement as the power-exit in our most optimistic scenariosgreen Covid recovery with 280 virtually global participation (3p). Figure 5 compares the PE trajectories of demand-3p-G and power-281 3p-G against NPi-green, NDC-green, and 1.5 o C to visualise their aggregate effects and illuminate the 282 remaining transformations necessary. The most glaring divergence between NPi-G and 1.5 o C pathways 283 is the 17-fold difference in non-electric coal consumption, which the power-exit further exacerbates. 284 Figure 5 suggests that natural gas restrictions and bioenergy support are the most urgent priorities 285 after coal, given the sharp, immediate bifurcation between their 1.5 o C trajectories and all other 286 pathways. Moreover, demand-3p-G incentivises an additional 780EJ gas and 2100EJ oil ( Figure SF2), 287 which can be avoided with immediate and sustained investment in renewable industry and transport 288 fuels. 289

290
Our three data-driven scenarios of post-Covid infrastructure (Appendix I) span a range of 1670GW-291 2320GW of coal power capacity in 2025 iii . DPE demonstrates the path-dependence of PPCA expansion 292 to these near-term uncertainties. Most notably, China accedes in neutral-2p (1070GW national 2025 293 capacity) and green-2p (980GW) scenarios but abstains in brown-2p (1310GW). Figure 3 illustrates the 294 dynamic impacts of China's decision while Figure 4 shows the disparities in long-term prospects. 295 We report coal efficacy indices (1p-3p range) of .29 (.03-.76) for demand-brown and .53 (.06-.86) for 296 demand-green, and mitigation efficacy scores of .12 (.01-.33) and .23 (.02-.38), respectively. Power-297 exit scenarios exhibit minimal overall sensitivity all analysis dimensions, meanwhile, with coal efficacy 298 scores ranging between -.03 (brown-1p) and .02 (green-3p), and mitigation efficacies between -.01 and 299 .01. Nevertheless, these results suggests a robust negative correlation between near-term coal power 300 capacity and long-term PPCA efficacy. Greener public investment and regulatory decisions at this 301 critical juncture not only reduce immediate emissions but also have legacy effects that facilitate future 302 feasibility of coal phase-out policies. Myopic brown recovery packages, meanwhile, would impose 303 substantial strain upon future generations to mobilise the necessary transition. However, coal power phases out in these scenarios amidst rapid coal and emissions declines economy-327 wide. The power-sector bias, evident throughout the coal phase-out discourse 24,28,31 , may be explained 328 in part by data accessibility barriers. The only open-access, comprehensive, coal-asset-level datasets iv 329 were power-plant-specific 48 until comparable data on mines v and steel plants vi were published in 2021. 330 We therefore surmise that the PPCA's sector-exclusivity was motivated by politicse.g. to encourage 331 maximum participationand by under-contextualised scientific messaging. 332 The inadequacy and short-sightedness of the verbatim PPCA is evidenced by the future coal demand 333 profile in REMIND's NPi scenarios; while electricity accounted for ~60% of 2015 coal use 49 , it represents 334 just 16% cumulatively from 2020-2100 ( Figure 5). Moreover, the power-exit generally decreases free-335 rider coal electricity while CtL and industrial coal use universally increase. Other model baselines 336 robustly corroborate coal demand growth in industry 50 and transport 51 . A recent review suggested that 337 model scenarios are often overly-dependent on coal, but some power sector bias was evident and it 338 found that coal phases-out most readily in REMIND's CEA simulations 24 . The present study does not 339 dispute the urgency of power sector decarbonisation, as electrification is vital to myriad mitigation 340 strategies 52 , but provides grounds for the coalition-of-the-willing to explicitly cover non-electric 341 sectors. 342 The Demanding Demand-exit 343 We acknowledge that COALogit cannot accurately estimate demand-exit feasibility since power-exit 344 PPCA pledges form our empirical basis. Our analysis assumes perfect interchangeability to directly 345 compare the two policy options, but a real-world trade-off is anticipated between policy ambition and 346 coalition growth. Stated political ambition, as insinuated by the first-round NDCs, supports this theory. 347 Relative to 1.5 o C-consistent levels, the NDC scenario leaves 10x as much residual non-electric coal use 348 as unabated coal power, which is phased-out faster than any PPCA scenario modeled here ( Figure 5). The Policy Feedback Loop

357
The evolving coalitions derived by COALogit are largely insensitive to policy choice, i.e. for a given Covid 358 recovery, power-exit and demand-exit coalitions are nearly indistinguishable. This is an artefact of 359 COALogit's parsimonious dependence on coal-power-shares and the fact that the power-exit is simply 360 a subdivision of the demand-exit. Generally, we observe an inverse relationship between OECD 361 coalition size and non-OECD accession probabilities due to extra-coalition leakage of coal electricity, 362 best illustrated by Figure SF1b-d. 363 Although demand-2p scenarios trigger net-negative extra-coalition coal leakage, free-rider coal power 364 consumption actually increases, discouraging their accession. Power-2p scenarios are also unique, in 365 that extra-coalition coal-fired electricity decreases. However, the root cause is a hindrance of end-use 366 electrification globally, notably exacerbating liquid-fueled transport, the most notoriously challenging 367 end-use to decarbonise across IAM scenarios 53 . Hence, PPCA members must counteract the negative 368 feedbacks provoked by their demand-side efforts and mobilise self-perpetuating policy uptake by 369 ramping up electrification, VRE, and knowledge transfer to maximise technological spillovers. 370 A Supplementary Supply-exit 371 Furthermore, recent literature highlights the importance of complementing demand-side policies with 372 supply-side action 54-56 through for example mining or export restrictions. This counteracts price 373 depression and leakage, increasing the potential for self-propagation. Given bilateral trade 374 partnerships and spatial variance in coal quality, however, policy efficacy depends upon the specific 375 adopters. 376 Crucially, the largest anticipated coal consumers in 2045 -China, India, and ASEAN members ( Figure  377 2c)can each sustain a self-sufficient coal supply. However, their coal infrastructure receives 378 significant overseas financing from OECD-based investors 57 , where divestment campaigns are 379 historically commonplace 56 . Granted, Chinese banks are the world's largest overall coal financiers 57 380 and may insulate the domestic industry from foreign politics, but OECD legislatures can conceivably 381 induce coal declines through cross-border financial mechanisms, e.g. debt-for-nature swaps 58 . China's 382 historical 22-year mean plant lifetime (Table 2) and its 2060 carbon neutrality pledge 40 breed cautious 383 optimism. 384 Averting the Next Crisis 385 These coal-rich nations also exhibit the highest path-dependence of accession probability to near-term 386 investment decisions. Most glaringly, China falls below the 2p threshold viii and Indonesia below 3p 387 probability in brown recovery scenarios. Additionally, we observe that numerous highly-probable 388 coalition members within the OECD continue to commission coal power plants in brown and neutral 389 Covid recoveries ix . PPCA accession then forces a sudden mass exodus of unamortised capitala 100% 390 rate of early retirement x from 2025-2030. Thus, to protect the health of their economy 28 , power grid 59 , 391 citizenry 23 , and global-leader status, OECD governments must cancel their entire coal pipelines 60 . 392

393
DPE presents a way forward for inter-disciplinary climate policy research seeking to understand the 394 intersection of techno-economic, socio-political, and climate target feasibility. To enable similarly 395 evidence-based simulations of policy uptake in future studies, empirical research must identify robust 396 correlations between revealed ambition, viz. domestic legislation, and energy-economic variables 397 computed endogenously by forward-looking models. As the remaining window to respect the 1.5 o C 398 target dwindles 2 , we invite the data science community to contribute their expertise in large-scale 399 regression exercises 16