Climate Forecasts Against the Record: Temperature, Emissions, Concentrations and the Energy Mix

Bottom line

Climate projections did well where they were physics and poorly where they were economics. Warming projections going back to Hansen et al. (1988) and the IPCC’s First Assessment Report land close to the record once each is scored on the radiative forcing that actually occurred. Atmospheric CO2 sits within 1.3% of the RCP8.5, RCP4.5 and RCP2.6 paths. Emissions were the weak link. They ran above even the highest IPCC scenarios through the early part of the last decade, then fell well below them, so that RCP8.5 now overstates the latest year by 21%. The energy-mix projections missed solar and wind by multiples, yet the fossil share of total primary energy came in within 3.3pp of every IEA projection we scored. The misses have layered causes that change over time. Before the turn, slower growth and falling energy intensity did most of the work. Since then, falling energy intensity has still done the most to cut emissions. A falling carbon intensity of energy, led by cheap wind and solar, has become narrowly the largest share of the gap against the early-century trend. The cleanest reading of the “policy versus technology” question is that it is not a clean split. Explicit carbon pricing is probably small, though we could not size it. Deployment and industrial policy in Europe, China and elsewhere helped drive the learning curves that made renewables cheap. The biggest caveat is the last decade of temperature. Warming has run at the upper edge of the assessed range, and whether that reflects aerosol cleanup or a higher climate sensitivity is unresolved.

Key judgments

The scorecard

The table below compares what was projected with what happened, variable by variable. It pools projections by how far ahead they looked, measured from publication. Each cell is the mean absolute percent error between projection and outcome, so a smaller number is a better forecast.¹

Three patterns stand out. Totals were projected better than their parts: total electricity and the fossil share of energy had the smallest errors, while solar, wind and coal had the largest. The physical chain from emissions to concentrations was projected better than the economic chain that sets emissions. And the warming errors look larger than the pre-specified hypothesis implied, partly because our test is strict and rests on a handful of projections and short windows.

Two cautions apply throughout. None of these were forecasts in the ordinary sense. The IPCC scenarios were built to span possible futures, with RCP8.5 deliberately near the top. The IEA’s reference cases project only enacted policy, so later policy shows up as a “miss” by design. Scoring them as forecasts is still useful, because people used them that way.

Temperature: the physics held up, with a hot last decade

The old projections, scored two ways

The canonical test is Hausfather, Drake, Abbott and Schmidt (2020), which compared a long series of past model projections with observed warming.² Its central move is to score each projection twice. As issued, a projection is judged on the warming it predicted, which mixes two kinds of error: wrong assumptions about emissions and forcing, and wrong climate physics. Scored per unit of forcing, it is judged on whether it got warming right given the forcing that actually occurred. Only the second is a test of the physics.

The paper found most of its projections consistent with observations within their uncertainty, particularly once forcing mismatches were accounted for. We re-scored its published data for the most prominent of them, Hansen’s and the IPCC’s central cases. Our test is stricter: whether each came within a quarter of observed warming. As issued, Hansen 1988 Scenario B projected 1.52 times the warming observed over 1988-2017 (trend ratio, Hausfather et al. 2020 data). Per unit of the forcing that actually occurred, Hansen 1988 Scenario B implied 1.20 times the observed warming response over 1988-2017. The IPCC’s First Assessment Report (IPCC 1990) is the other prominent case. As issued, IPCC FAR 1990 (best) projected 1.37 times the warming observed over 1990-2017 (trend ratio, Hausfather et al. 2020 data). Per unit of the forcing that actually occurred, IPCC FAR 1990 (best) implied 0.88 times the observed warming response over 1990-2017. Overall, 2 of the 8 projections came within a quarter of observed warming as issued, and 4 did so once scored on realized forcing.

The reason the early projections ran warm is well documented. Hansen et al. (1988) ran the GISS Model II, whose equilibrium sensitivity was 4.2 °C for doubled CO2.³ The IPCC’s sixth assessment puts the outcome plainly. IPCC AR6 WG1 assesses that Hansen et al. (1988) Scenario B projected about 50 percent more warming than observed over 1988-2017, largely because it overestimated subsequent radiative forcing.⁴ The FAR made the same kind of error. The observational estimate of the actual increase in anthropogenic forcing over 1990-2017 is 1.1 W/m² (Dessler and Forster 2018), against the FAR’s projected 1.6 W/m2. Under the forcings that actually occurred, the change in temperature projected by the IPCC FAR aligns with observations (Hausfather et al. 2020).

The AR6 verdict is the right summary, with its own caveat attached. AR6 assesses with medium confidence (limited evidence, high agreement) that past projections of global temperature are consistent with subsequent observations, especially once the difference between assumed and actual radiative forcing is accounted for. The same chapter warns against reading too much into the agreement. AR6 notes that agreement between past projections and observations could be fortuitous, the product of compensating errors such as too-low climate sensitivity combined with too-strong forcing.

Observed global warming against Hansen 1988 and IPCC projections

Start

Extending the test to the present

Hausfather and colleagues end their scoring windows several years before the present. We extended the as-issued comparison to the latest full year, without the forcing adjustment, which would need forcing series not assembled here. The result is less flattering to the newer projections than to the older ones. Extended to 2025, IPCC FAR 1990 (best) projected 1.16 times the observed warming trend over 1990-2025 (as issued, not forcing-adjusted). Extended to 2025, IPCC TAR 2001 (best) projected 0.69 times the observed warming trend over 2001-2025 (as issued, not forcing-adjusted). Extended to 2025, IPCC AR4 2007 (multi-model mean) projected 0.66 times the observed warming trend over 2007-2025 (as issued, not forcing-adjusted). The TAR and AR4 central projections have so far under-projected warming by about a third. The FAR’s central case is now close to the record.

The most recent decade explains most of that. The observed warming trend over 2015-2025 was 0.29 °C per decade (mean of three records, linear fit). That pace is well above the long-run rate: The observed warming trend over 1988-2025 was 0.22 °C per decade (mean of three records). In the AR4 low-to-high series that Hausfather and colleagues digitized, the low end projected 0.24 °C per decade and the high end 0.26 °C for the same period. The observed trend therefore sits above that envelope; on a scale where the low end is zero and the high end is one, it sits at 2.02. That envelope is narrow, however, and it is not the assessed likely range. The comparison the IPCC itself would make, against the AR6 near-term range, comes out inside the range. Observed 2015–2024 warming is broadly consistent with AR6’s projected 2001–2020 to 2021–2040 warming rates under SSP2-4.5, whose very likely range tops out at 0.036 °C per year.⁵ The latest update says the same about the latest decade. The observed warming rate falls within the upper end of AR6’s projected 2001-2020 to 2021-2040 range under SSP2-4.5 (0.016 to 0.036 °C per year).⁶

So the last decade ran hot against the TAR and AR4 central cases but inside AR6’s assessed range, at its warm edge.

Has warming accelerated?

The live dispute is whether the warm edge is a signal. Both sides agree on the facts of the record. The decade 2016-2025 was 0.32 °C warmer than 2006-2015, above the AR6-assessed 1980-2020 rate of 0.19 °C per decade. The rate of human-induced warming over 2016-2025 is 0.27 °C per decade (likely range 0.2-0.4), versus 0.2 °C per decade quoted in AR6 for 2010-2019. What they disagree about is cause, and therefore about what comes next.

The Forster group, which produces the annual indicators update for the IPCC community, reads the rise as largely explained. Of the rise from AR6’s 0.2 °C per decade to 0.27, only 0.01 °C per decade is a real increase in rate; the rest is rounding and method/data updates. The update attributes the acceleration mainly to aerosols. Falling aerosol emissions have been the predominant driver of the acceleration in anthropogenic warming since 2000-2009, and a slowing in aerosol declines is producing a slight net deceleration over the last three years. Among the aerosol changes, the cleanup of ship fuel is the best measured. Sulfur dioxide from international shipping fell from 10.4 to 2.0 Tg in 2020, a drop of 8.4 Tg after the IMO fuel rule. Falling shipping sulfur emissions since 2020 weakened aerosol cooling, adding about +0.1 W/m² of forcing over 2020–2024.

Hansen and coauthors read the same record as evidence that the IPCC has underestimated both aerosol cooling and climate sensitivity.⁷ Hansen et al. estimate that warming since 2010 has run more than 50 percent faster than the 1970-2010 rate of 0.18 C per decade. Hansen et al. infer a ship-aerosol forcing of about 0.5 W/m² from the 2020 sulfur-fuel rule, an order of magnitude above the IPCC-based estimate. Their earlier paper made a forecast that can now be scored.⁸ Hansen et al. predict that, with current policies, post-2010 global warming will run at least 0.27 C per decade, reaching 1.5C in the 2020s and 2C before 2050. The upper edge of Hansen et al.’s predicted post-2010 accelerated warming rate is 0.36 C per decade. On the observed trend of 0.29 °C per decade over the latest eleven years, that forecast is holding. The same group concludes that sensitivity is high. From the glacial-to-interglacial record, Hansen et al. infer equilibrium climate sensitivity of 4.8 °C for doubled CO2 (1.2 +/- 0.3C per W/m2), above the IPCC AR6 best estimate of 3C.

The ship-aerosol number is where the two views part, and the spread is wide. The change in aerosol forcing from the 2020 ship-fuel limit, calculated with IPCC’s formulation, is 0.079 W/m². A review of five ship-aerosol modeling studies finds a mean forcing of 0.12 W/m² (range 0.07 to 0.15). Hansen’s 0.5 W/m² is several times the top of that range. The difference is not measurement error that more data will soon average away: Hansen et al. concede the key input is not observed directly. Hansen et al. concede modern temperature change does not tightly constrain climate sensitivity because aerosol forcing is not measured.

Hansen’s most specific near-term call has held on its own terms. Hansen et al. predicted global temperature would remain above +1.5 C at the end of 2024 and at most barely fall below +1.5 C in 2025. On his 1880-1920 baseline, GISTEMP puts 2025 at 1.45 °C, just 0.05 °C short of the line, and the 2023-2025 average at 1.48 °C. Against 1850-1900 the Forster update gives 1.39 °C and our three-record mean 1.39 °C; the apparent miss there is the baseline, not the forecast. A call about one hot spell is a weak test, though, and it does not settle the sensitivity question.

Our view: the temperature scorecard is good, and its errors are dominated by input assumptions rather than physics. The aerosol and sensitivity dispute is real and cannot be settled from the global-mean record alone.

Emissions: too low, then too high

The early miss: emissions ran above the top scenario

At the turn of the century the IPCC’s Special Report on Emissions Scenarios (IPCC 2000) set out its marker paths, from the coal-heavy A1FI to the low-growth B1. Within a few years the world had outrun most of them. Raupach et al. (2007) documented the acceleration.⁹ Global fossil-fuel and industrial CO2 emissions grew 3.2 percent per year in 2000-2004 (EIA data), roughly triple the 1990s rate. Observed global fossil CO2 emissions after 2000 tracked close to the highest SRES scenario (A1FI) and grew faster than it. The cause was not faster population or income growth. The post-2000 acceleration came from a halt or reversal in the declines of the carbon intensity of energy and the energy intensity of GDP, not from faster population or income growth. The mechanism was concentrated in one country. China’s declining energy intensity of GDP reversed around 2002, and together with very fast income growth drove its rapid emissions growth in the early 2000s.

Raupach and coauthors also identified the assumption that failed. All IPCC scenarios postulate continuous declines in energy intensity and carbon intensity; their stalling after 2000 is why even the fossil-intensive scenarios underestimated emissions growth. The scenario authors replied that the SRES paths were never meant to be tracked year by year.¹⁰ That is right as a matter of design, but it does not change the record. Our own numbers confirm the early miss. SRES A1FI projected 2010 fossil CO2 emissions -3% away from the outcome.

Peters et al. (2013) extended the finding a few years further, concluding that emissions continued to track the high end of the scenarios.¹¹

The turn

The growth rate then fell sharply. Global fossil CO2 grew 2.4% a year over 2000-2013. Global fossil CO2 grew 0.8% a year over 2013-2024. The Global Carbon Budget, using its own periods, shows the same shift.¹² Global fossil CO2 emissions grew 2.8 percent per year in the 2000s, then 1.2 percent in the 2010s and 0.8 percent over 2015-2024.

Global fossil CO2 emissions against the SRES and RCP scenarios

By the latest observed year, the outcome sat in the middle of the scenario range. Global fossil CO2 emissions were 37.8 Gt of CO2 in 2024 (Global Carbon Budget 2025, net of cement carbonation). Against that outcome, the high scenarios all overshot. SRES A1FI projected 2024 fossil CO2 emissions 22% away from the outcome (positive = above). SRES A2 projected 2024 fossil CO2 emissions 17% away from the outcome (positive = above). SRES A1B projected 2024 fossil CO2 emissions 25% away from the outcome (positive = above). RCP8.5 projected 2024 fossil CO2 emissions 21% away from the outcome (positive = above). The middle and low scenarios bracketed it. RCP4.5 projected 2024 fossil CO2 emissions 0% away from the outcome (positive = above). SRES B1 projected 2024 fossil CO2 emissions 2% away from the outcome (positive = above). The newer SSPs, which began from a later base year, show the same pattern. SSP5-8.5 projected 2024 fossil CO2 17% away from the outcome, nine years after its 2015 base. SSP2-4.5 projected 2024 fossil CO2 2% away from the outcome.

The hypothesis we set going in was that emissions would land below the A1B, A2 and RCP8.5 paths and above B1 and RCP2.6. It held except at B1, which sits 2% above the outcome and so missed narrowly on the wrong side. Observed emissions ran at or above A1FI until the early part of the last decade, and above A2 a year longer. The sign of the miss flipped around the time global coal use is thought to have peaked.¹³

The RCP8.5 argument, resolved by the timing

This timing reconciles a literature that looks contradictory. Schwalm, Glendon and Duffy (2020) argued that RCP8.5 remained the best guide.¹⁴ Among the RCPs, RCP8.5 matched historical total cumulative CO2 emissions for 2005-2020 most closely, within 1 percent. Hausfather and Peters (2020) replied that the match was an accident of two offsetting errors.¹⁵ Schwalm et al.’s apparent match between near-term cumulative emissions and RCP8.5 is attributed to compensating errors: too-high fossil CO2 and too-low land-use emissions. The land-use numbers bear them out. The RCPs assume land-use emissions as low as 1.5 Gt of CO2. The Global Carbon Project’s current best estimate of land-use-change emissions is 5.5 Gt of CO2 (plus or minus 2.6).

Our own numbers show the same split. RCP8.5 projected cumulative 2005-2024 fossil CO2 8.9% away from the outcome. Counting land-use emissions as well, RCP8.5’s cumulative 2005-2024 total CO2 was 2.7% away from the outcome. RCP8.5 looks right on total cumulative CO2 because it assumed land-use emissions far below the current estimates, while its fossil emissions ran well above the outcome. The near-term case for RCP8.5 as business as usual rested on the early period, and the later period has undercut it.

Pielke, Burgess and Ritchie (2022) put the point in its most systematic form.¹⁶ Only 6 percent of AR5 scenarios matched observed 2005-2020 fossil-and-industry CO2 growth to within 0.1 percent a year. Every baseline (‘business as usual’) scenario in the AR5 and SSP databases over-projected 2005-2050 fossil emissions growth by more than 0.1 percent per year relative to observations and IEA STEPS. They found the plausible scenarios cluster lower. The scenarios consistent with observed and IEA-projected emissions imply a median of 2.2 °C of warming by 2100. One caveat in their own results is important. Screening on 2005-2019 data alone (before COVID) leaves upper limits of roughly 70-100 GtCO2 a year in 2100, so the exclusion of >3 °C paths depends on the IEA projection. The exclusion of the hottest paths depends in part on trusting the IEA’s projection of the next few decades, a projection whose track record is the subject of this briefing.

Where the shortfall came from geographically

The emissions miss is mostly a China story, in both directions. Over the early period China drove the acceleration. China’s emissions rose by 6.3 Gt of CO2 over 2000-2013. World emissions rose by 9.8 Gt of CO2 over 2000-2013. Since then its growth has slowed sharply but not stopped. China’s emissions rose by 2.35 Gt of CO2 over 2013-2024.

To size the later shortfall we compared each emitter’s latest-year emissions with a straight-line continuation of its own trend over the high-growth years. Against a continuation of its 2000-2013 trend, world fossil CO2 in 2024 fell short by -4.9 Gt of CO2. China accounted for -3.0 Gt of CO2 of that 2024 shortfall against the 2000-2013 trend. India is the exception, running above its own trend. India’s 2024 emissions exceeded the continuation of its own 2000-2013 trend by 0.34 Gt of CO2 (negative = below). The United States and the EU fell in absolute terms. U.S. emissions changed by -0.57 Gt of CO2 over 2013-2024. EU27 emissions changed by -0.75 Gt of CO2 over 2013-2024.

Where world emissions fell short of their 2000-2013 trend, 2024

Of the pre-specified hypothesis, the China part holds, and the India part holds against India’s own trend. Regional scenario paths were not available to score India, the United States or the EU directly; the U.S. and EU declines are real but small in the global total.

Concentrations: the carbon cycle behaved roughly as modeled

Level close, increment less so

Atmospheric CO2 is the variable the scenarios got most nearly right. Global mean CO2 was 422.8 ppm in 2024 (NOAA marine-surface network). Global mean CO2 was 425.6 ppm in 2025. RCP8.5 projected 2024 CO2 concentration 1.3% away from the observed level.

Atmospheric CO2 against the RCP and SSP concentration paths

A level built up over two centuries is an easy target. The harder test is the rise since each scenario’s start, and there the errors are larger. Measured on the rise since 2005, RCP8.5 projected 12% more (or less, if negative) CO2 build-up through 2024 than occurred. RCP8.5’s concentration miss is in the same direction as its emissions miss, and smaller. That is what one would expect if the sinks behaved about as the models assumed.

The airborne fraction moved more than the hypothesis allowed

The pre-specified hypothesis held that the airborne fraction stayed roughly constant decade by decade. It held only loosely. The airborne fraction of CO2 averaged 36% in the 1960s (atmospheric growth over fossil plus land-use emissions, Global Carbon Budget). The airborne fraction of CO2 averaged 46% in the 2010s (atmospheric growth over fossil plus land-use emissions, Global Carbon Budget). The airborne fraction of CO2 averaged 51% in the 2020-2024 (atmospheric growth over fossil plus land-use emissions, Global Carbon Budget).

Share of CO2 emissions that stayed in the air, by decade

The last of those numbers is inflated by a single event. The atmospheric CO2 growth rate in 2024 was a record 3.73 ppm per year, mainly due to the 2023/2024 El Nino. The Global Carbon Budget also records that the land sink has weakened. Climate change and variability reduced the land CO2 sink by 25 percent on average over 2015-2024 (and the ocean sink by 7.1 percent). The land sink took up 21 percent of total emissions over 2015-2024 and has been relatively stagnant since 2000. Over its full record the Budget gives a lower and steadier figure than our decadal series. Over 1959-2024, 44 percent of total anthropogenic CO2 emissions (fossil plus land use) remained in the atmosphere. The decadal drift also depends on the land-use term in the denominator. The Global Carbon Budget describes its land-use emissions estimates as having large uncertainty and limited reliability.

Even the record-growth figure depends on the measure. The source put 2024 CO2 growth at about 3.7 ppm; the difference of NOAA GML global annual means, 2023 to 2024, is 3.44 ppm, the largest since 1980 but only narrowly (exploratory check). Measured as Global Carbon Budget atmospheric growth converted at 2.124 GtC per ppm, 2024 growth was 3.73 ppm, in line with the source’s 3.7 ppm (exploratory check). The two answer slightly different questions, and the difference decides whether the year is a clear record.

Our reading is that the carbon cycle behaved close to what the scenarios assumed. The drift toward a higher airborne fraction is exaggerated by the latest El Niño and may partly reflect the uncertain land-use estimate.

The non-CO2 misses

Methane departed further from the scenarios than CO2 did. A plateau early in the century gave way to renewed growth. From 1999 to 2006 atmospheric methane rose only 0.6 ppb/yr, a near-stabilization. Atmospheric methane growth averaged 7.6 ppb/yr in 2010-2019. Atmospheric methane growth reached a record 18.0 ppb/yr in 2021. The methane budget now places observed emissions near the top of the scenario range.¹⁷ Since 2012, global direct anthropogenic methane emissions have tracked the no-or-minimal-mitigation scenarios SSP5 and SSP3. The mitigation-scenario reductions in methane (SSP1 and SSP2) have not happened. Part of the rise is a climate feedback rather than human emissions. Post-AR6 literature finds wetland methane emissions rose 6-7 Tg CH4 per year in 2010-2019 versus 2000-2009, mostly from temperature-driven climate change, a feedback not counted in anthropogenic emissions.

Aerosols are the other non-CO2 miss, and they cut the other way. Revised inventories show China’s sulfur cleanup came faster than the models assumed. Revised CEDS data cut China’s SO2, black and organic carbon emissions in 2014 by about 50 percent relative to the data used for CMIP6 and AR6. Cleaner air reduced the cooling that aerosols had been providing. Radiative forcing rose 0.38 W/m² (10 percent) in the six years since AR6, from rising GHG concentrations combined with weaker aerosol cooling.

For the temperature scorecard, both misses push the same way: methane forcing came in higher than the mitigation scenarios assumed, and aerosol cooling weaker. The CO2 forcing that dominates came in near the central paths. That combination is consistent with a last decade at the warm edge of the assessed range without requiring a change in sensitivity.

The energy mix: missed by multiples inside electricity, close on the total

Solar and wind outran every IEA reference case

The IEA’s World Energy Outlook is a standard benchmark for energy projections, and its renewables record is poor. Our scoring of seven editions’ reference and current-policies cases against Ember’s generation data shows the scale.¹ Realized 2020 world solar generation was 31.57 times what the 2002 WEO Reference Scenario projected. Realized 2020 solar generation was 5.46 times the 2010 WEO Current Policies projection. Realized 2020 solar generation was 2.66 times the 2012 WEO Current Policies projection. Realized 2020 solar generation was 1.81 times the 2015 WEO Current Policies projection. Realized 2025 solar generation was 13.96 times the 2008 WEO Reference projection. World solar generation reached 2,807 TWh in 2025 (Ember). The 2008 WEO Reference Scenario had projected 201 TWh of solar for 2025.

World solar generation: IEA WEO projections against the outcome

Start

Wind was missed by less, but still by large factors in the earlier editions. Realized 2020 wind generation was 5.18 times the 2002 WEO Reference projection. Realized 2020 wind generation was 1.47 times the 2010 WEO Current Policies projection. World wind generation reached 2,705 TWh in 2025 (Ember).

World wind generation: IEA WEO projections against the outcome

Start

The hypothesis we set going in, that editions from the early part of the century underprojected solar by more than five times and wind by more than two, holds for those editions. It does not hold for the last editions scored: the WEO-2012 and WEO-2015 solar misses were 2.66 and 1.81, and the WEO-2012 wind miss was 1.39. The IEA learned, slowly.

The academic literature found the same pattern before the most recent data arrived. Metayer, Breyer and Fell (2015) traced the IEA’s renewables projections edition by edition.¹⁸ Metayer, Breyer and Fell find actual installed solar PV capacity in 2014 was about 3 times the WEO 2010 projection for that year. WEO 2010’s solar PV capacity projection for 2020 was already achieved in 2012, two years after publication. WEO 2002’s 2030 wind capacity projection was reached in 2010, 20 years early. Their explanation is simple. The paper attributes the WEO renewables misses to an assumption of linear rather than exponential growth.

Coal and nuclear ran below projections

The other side of the electricity miss is fossil and nuclear generation that the IEA overprojected. Realized 2025 coal generation was 0.78 times the 2008 WEO Reference projection. Realized 2020 coal generation was 0.81 times the 2010 WEO Current Policies projection. Realized 2025 nuclear generation was 0.83 times the 2008 WEO Reference projection. Realized 2020 nuclear generation was 0.75 times the 2010 WEO Current Policies projection. Gas went the other way. Realized 2025 gas generation was 1.18 times the 2008 WEO Reference projection. The coal pattern has a history. In the early part of the century the IEA underprojected coal, when China’s buildout ran ahead of the outlooks. Early-2000s WEOs underprojected 2010 coal power capacity by an average of about 350 GW, or 21 percent. Since then it has overprojected. Nuclear has been overprojected throughout. WEO reports of the 2010s projected nuclear capacity rising about 50 percent by 2035 from 2012 despite stagnant capacity.

The total held up

Despite those misses, total electricity was projected well. Realized 2025 total generation was 1.05 times the 2008 WEO Reference projection. The fossil share of total primary energy, the variable that matters most for emissions, was projected well too. The realized fossil share of world primary energy in 2024 was 79.4% (IEA basis, WEO-2025 Table A.1a). Across the seven WEO vintages (2002-2015), the projected 2024 fossil share missed the outcome by at most 3.3pp. On average the vintages projected a 2024 fossil share 1.5pp above the outcome.

Fossil share of world primary energy in 2024: WEO projections against the outcome

This is the most important single fact about the energy-mix scorecard. The IEA’s own assessment agrees. Fossil fuels supplied nearly four-fifths of global energy demand in 2024, a share the IEA says has fallen only marginally since 2000.¹⁹ Solar and wind grew from a small base, and a projection that missed them by a multiple still missed the total energy mix by little. The surprise sits inside electricity, which is a minority of final energy, and it has so far mostly displaced growth in coal rather than cut deeply into the stock of fossil energy.

Why the IEA missed: costs, conventions, or conservatism

The literature offers three complementary explanations.

The first is costs. Way, Ives, Mealy and Farmer (2022) scored thousands of model projections of solar cost declines.²⁰ Way et al. examined 2,905 integrated assessment model scenarios’ projections of the annual rate of solar PV system cost decline between 2010 and 2020. The mean integrated-assessment-model projection was for solar PV system costs to fall 2.6 percent per year between 2010 and 2020, and every projection was below 6 percent. Solar PV costs actually fell about 15 percent per year between 2010 and 2020. Past projections of present renewable energy costs by influential energy-economy models have consistently been much too high. The trend they fit is simple. For several decades the costs of solar PV, wind and batteries have dropped roughly exponentially at a rate near 10 percent per year. Projections that used that trend would have done better than the models did.

The second is the convention of projecting only enacted policy. The IEA’s current-policies case assumes no policy beyond what is in law. Feed-in tariffs, auctions and Chinese industrial programs that came later therefore show up as misses. Mohn (2020) weighs this and rejects stakeholder bias.²¹ Mohn rejects stakeholder (member-state and oil-industry) bias as the explanation and favors institutional conservatism, capital-vintage effects and adjustment costs. Mohn judges that the IEA’s under-projection of solar and wind is more likely due to biased parameterization, restrictions and cost assumptions than to the functional form of the World Energy Model. His most useful observation cuts against the usual critique. Ex post, the IEA’s Current Policies scenario was the most accurate guide to aggregate energy demand and greenhouse gas emissions, while new renewables tracked the 450 scenario. The current-policies case got the aggregates right while missing the composition.

The third is how models represent technology. In the WEO, technological change is exogenous and continuous, so technology shocks or sudden breakthroughs are ruled out. Jaxa-Rozen and Trutnevyte (2021) find that the modeler matters more than the cost input.²² General scenario features alone (organization, model, date, policy framing) classify scenarios into PV-growth quintiles with 73 percent accuracy. Adding PV capital-cost assumptions raises classification accuracy only to 77 percent for the 116 scenarios that report costs, and cost ranks below date, archetype, model boundaries and constraints, and climate policy as a predictor. The IPCC AR5 and SR1.5 scenario ensembles are relatively conservative about PV growth and capital costs, particularly in the near term.

Our view: the solar miss was mainly a learning-rate miss, compounded by a policy convention that excluded the deployment policies that drove the learning. The two are hard to separate because policy-driven deployment is what moves a technology down its learning curve.

Why it ended up different: growth, efficiency and carbon intensity

The Kaya accounting

Emissions are the product of four factors: population, income per person, energy per unit of income, and CO2 per unit of energy. A decomposition apportions any change in emissions among them. It is an accounting identity, not a causal model: it says which factor moved, not why.

We ran two versions: one decomposes the actual change in world emissions, the other the gap between the latest year and a path on which each factor kept its early-century growth.

On the second measure the four factors contributed in similar measure. World fossil CO2 in 2024 was -7.81 Gt of CO2 relative to a path on which all four Kaya factors kept their 2000-2013 growth rates. Slower growth of GDP per person accounted for -1.05 Gt of CO2 of that gap. Slower population growth accounted for -1.09 Gt of CO2. Faster decline in energy per unit of GDP accounted for -2.74 Gt of CO2. Faster decline in CO2 per unit of energy accounted for -2.93 Gt of CO2. Slower growth of population and income, the “other factors” in the question, account for a little over a quarter of the gap. Faster improvement in energy intensity accounts for about a third, and a falling carbon intensity of energy for the largest single share, narrowly.

The timing matters more than the total. On the first measure, carbon intensity reversed sign between the two periods. Over 2005-2013 the carbon-intensity term was 0.87 Gt of CO2. Over 2013-2024 the carbon-intensity term was -1.82 Gt of CO2. Over 2005-2013 the energy-intensity term was -4.08 Gt of CO2. Over 2013-2024 the energy-intensity term was -7.28 Gt of CO2. Before the turn, carbon intensity was rising, as China’s coal buildout added carbon-heavy energy, and energy intensity was the only factor holding emissions back. Since the turn, carbon intensity has fallen, though energy intensity is still the larger term in absolute size.

That qualifies the pre-specified hypothesis. Before the turn, slower growth and falling energy intensity explain essentially all of the shortfall. Since then carbon intensity has become a large term, narrowly the largest on the trend-gap measure, but energy intensity still dominates the actual decline.

China’s slowdown was mostly growth

China dominates the global numbers, and its gap is mostly about growth. China’s 2024 emissions were -10.96 Gt of CO2 relative to its 2000-2013 Kaya-factor trends. Slower GDP-per-person growth accounted for -6.68 Gt of CO2 of China’s gap. The China gap is larger than the world gap because continuing China’s early-century growth rates for another decade produces a very high counterfactual; the shares are what matter. The Global Carbon Budget reaches the same conclusion. China’s fossil emissions growth fell from 6.7 percent a year in 2005-2014 to 2.5 percent in 2015-2024, driven by weaker economic growth (about 5 percentage points), partly offset by a slower decline in energy per GDP (about four points). The IEA credits structural change as well as growth. The IEA dates the peak of China’s industrial coal consumption to 2014 and credits Chinese economic restructuring with the marked slowdown in global industrial energy-use growth over the past decade.

This is the main argument for treating part of the post-turn shortfall as cyclical. Slower Chinese growth reflects in part the end of an investment-led boom, not climate policy, and a growth recovery would reverse some of this term. The structural case rests on the latest data. The Global Carbon Budget projects that China’s fossil CO2 emissions rose only 0.4% in 2025, a slight increase. UNEP’s updated current-policy projections show China’s emissions peaking around 2025 and then falling, where earlier projections had growth to 2030. Which reading is right is open.

The developed-country decline

Le Quéré et al. (2019) studied the developed economies whose emissions peaked and then declined.²³ Emissions in the 18 peak-and-decline countries fell by a median -2.4 percent per year a year over 2005-2015. A falling fossil share of final energy accounted for a median 47 percent of the 2005-2015 emissions decline in the 18 countries. Integrated assessment models assumed annual GDP growth of 2.4 percent per year, more than double the rate observed in the peak-and-decline countries. In these countries, slower growth again explains much of the gap against the scenarios. Burgess, Ritchie, Shapland and Pielke (2021) generalize the point.²⁴ The 2005-2017 divergence of CO2 emissions from IPCC baseline scenarios is primarily due to the baselines over-projecting per-capita GDP growth, with all SSP baselines too high in all regions. From 2005 to 2017, global energy CO2 emissions grew more slowly than in 83 percent of IPCC AR5 baseline scenarios. They also identify the modeling assumption that failed in the other direction. Most AR5 and SSP baselines assume carbon intensity would not decline without climate policy, whereas the past decade suggests factors beyond climate policy cause carbon-intensity declines. One reason high-emission baselines diverge on coal is that they substantially over-projected the cost of solar power relative to coal and other fossil fuels.

Burgess and coauthors also checked the baselines’ intensity assumptions. Globally, AR5 and SSP baseline energy-intensity projections were consistent with observed 2005-2017 trends. AR5 baseline scenarios slightly under-projected the decline in carbon intensity in every region over 2005-2017. The baselines got growth wrong, not intensity, consistent with our accounting for the early period.

Policy, technology or something else

What can be measured directly

We sized the two channels of carbon-intensity improvement that can be bounded from generation data.

U.S. coal-to-gas switching in power over 2005-2024 avoided about 0.46 Gt of CO2 in 2024 (gas growth that replaced coal, times the coal-gas emission-factor gap; Ember factors). That U.S. switch equals 6% of the world’s 2024 shortfall against its 2000-2013 trend. Shale gas made that switch possible, and the IEA describes shale as a joint product of energy policy and private enterprise. The IEA says the rise of U.S. shale was driven by energy policy together with private innovation, risk-taking capital markets and high global energy prices.

Wind and solar did far more. Wind and solar generation added worldwide since 2013 displaced at least 2.3 Gt of CO2 in 2024 (if it displaced gas) … … and at most 3.5 Gt of CO2 (if it displaced coal). That bound is 30% of the world’s 2024 shortfall against the 2000-2013 trend at the gas-displacement end … … and 45% at the coal-displacement end. Wind and solar displacement since 2013 equals 79% of the carbon-intensity part of that shortfall at the gas-displacement end (above 100 means other forces pushed carbon intensity the other way). The IEA’s own estimate, made a different way for a shorter period, is in the same range. The IEA estimates that low-emissions technologies deployed since 2019 (solar PV, wind, nuclear, electric cars and heat pumps) now avoid 2.6 Gt of CO2 a year, about 7 percent of global emissions.

Emissions avoided in 2024 by coal-to-gas switching and by wind and solar growth

So the carbon-intensity improvement since the turn is mostly wind and solar, and the U.S. coal-to-gas switch is a small piece of it globally. We could not size explicit carbon pricing on the same basis. No published estimate of emissions abated by the EU Emissions Trading System was among the sources, and we chose not to infer it as a residual. The hypothesis that explicit pricing is the smallest of the three channels is therefore a judgment from the literature here, not a measured result.

Climate legislation as a whole

The broadest estimate of policy comes from Eskander and Fankhauser (2020), who studied climate laws in a large panel of countries.²⁵ The analysis covers 1,092 mitigation laws in 133 countries over 1999-2016. Each new national climate mitigation law reduces annual CO2 emissions per unit of GDP by about 0.78 percent in the short term (first three years). Each new national climate mitigation law reduces annual CO2 emissions per unit of GDP by about 1.79 percent in the long term (beyond three years). By 2016, national climate laws were associated with annual global CO2 emissions 5.9 Gt of CO2 lower than a no-legislation counterfactual. The 2016 legislation effect equals a cut in annual CO2 emissions of about 15 percent from the no-legislation baseline. Globally, climate legislation offset only about a third of CO2 emissions growth between 1999 and 2016, while in OECD/EU countries it is credited with holding CO2 at roughly 1999 levels.

Two features matter for reconciling this with the accounting above. The outcome is CO2 per unit of GDP, so a law that improves efficiency counts. And the design removes the common global trend in technology. The global fall in clean technology costs is absorbed by the year fixed effect, so the law coefficients are estimated net of common technology trends. Cost declines common to all countries are absorbed by the year effect. The estimate measures policy’s effect on top of technology, not policy’s role in creating it.

Why technology unrelated to climate policy is mostly an empty category

The solar learning curve did not happen on its own. Grubb and coauthors reviewed the induced-innovation literature.²⁶ Decomposing solar PV cost declines, Kavlak et al. (2018), as reviewed, estimate that public R&D and spillovers accounted for almost 50 percent of cost reductions over 1980–2000. From 2001 to 2012, public R&D and spillovers accounted for roughly one quarter (25 percent) of solar PV cost reduction, with scale economies and learning-by-doing accounting for about half (Kavlak et al. 2018, as reviewed). Kavlak et al. (2018) suggest the remaining PV cost reduction, attributed to private R&D, was largely catalyzed by deployment-support policies such as feed-in tariffs. The review concludes that numerous lines of evidence point to dominant causality from at-scale deployment to cost reduction.

Those deployment policies were climate and energy policies, from European feed-in tariffs to Chinese support for manufacturing and deployment. The cost decline that the IEA projections missed was therefore induced in large part by policies that the IEA’s convention excluded until they were law. The review also finds that targeted policy pulls harder than carbon pricing. The review judges that well-designed targeted policies may provide a far stronger pull on innovation than any plausible level of general carbon pricing.

The clean categories of technology unrelated to climate policy are narrow. Shale gas is one, though even it had policy support. General efficiency progress is another, and part of it responds to prices rather than policy. Knittel (2012) finds gasoline prices, not fuel-economy standards, were the principal driver of a 60 percent improvement in U.S. car and truck fuel efficiency over 1990–2006, once weight and power are controlled for. The IEA itself says both forces operate. The IEA states that innovation and technology progress occur both in response to policies and independently of them.

A rough apportionment

Before the turn there was no shortfall to explain. Emissions ran at the top of the range, and the scenarios that did best happened to project slower growth and faster efficiency gains.

Since the turn, three forces of similar size explain why emissions fell below the high scenarios. Slower growth, concentrated in China and only partly structural, accounts for a little over a quarter of the gap against trend. Faster improvement in energy intensity accounts for about a third; its causes mix structural change in China’s economy, efficiency technology, prices and efficiency policy. Falling carbon intensity accounts for narrowly the largest share on this measure, and most of it is wind and solar displacing coal and gas.

On the policy question, then, the answer depends on how one counts. Explicit carbon pricing is probably the smallest channel, though we could not size it. Climate legislation in total, net of common technology trends, is a material but minority contributor. Deployment policy induced part of the cost decline behind the carbon-intensity term; the evidence covers solar more than wind or batteries. Even counting all wind and solar displacement as policy puts policy’s share (30% to 45%) about level with energy intensity’s, not clearly above it. Counting only the induced part puts it lower. Policy and technology are entangled, and any single ratio hides that.

What this means for today’s projections

The current benchmarks have moved down

Today’s benchmarks embed the lessons of the last two decades, and they have moved down accordingly. The IEA’s 2025 World Energy Outlook puts global energy-related CO2 emissions in 2050 under its Current Policies Scenario some 10 Gt lower than the same scenario projected when it was last modeled in 2019. The IEA now projects emissions flat to falling under existing policies. In the IEA’s Current Policies Scenario, global energy-related CO2 emissions rise from 38 Gt in 2024 to 39 Gt in the early 2030s and stay around that level to 2050. In the IEA’s Stated Policies Scenario, global energy-related CO2 emissions fall to around 35 Gt by 2035, a level last seen in 2015. The implied warming has come down with them. The IEA’s projected 2100 warming is almost 3 °C under its Current Policies Scenario and 2.5 °C under its Stated Policies Scenario. In its 2025 report, UNEP projected that a continuation of current policies would hold warming this century below 2.8 °C with a 66 percent chance.²⁷ UNEP’s newest edition, released in late October, was not among the sources read here.

The IPCC itself has downgraded the high scenarios. AR6 judges the likelihood of the high-emissions RCP8.5 and SSP5-8.5 scenarios to be low in light of recent energy-sector developments, and notes that IEA WEO 2020 stated policies implies roughly constant fossil CO2 emissions to 2070, in line with RCP4.5, RCP6.0 and SSP2-4.5. Pielke, Burgess and Ritchie (2022) found that the IEA’s projection lined up with the plausible scenarios. The median 2050 emissions of the plausible scenarios are essentially identical to the IEA’s independent WEO 2021 projection.

But the downward revisions have just reversed

The latest round of projections moved the other way for the first time. The STEPS warming outcome in the 2025 edition is about 0.1 °C higher than in the 2024 edition, the first year the IEA reports the gap between stated policies and climate targets widening rather than narrowing. The IEA’s 2025 Stated Policies Scenario has the United States with 30% less renewables capacity in 2035 than the 2024 edition did. The 2025 Stated Policies Scenario projects coal demand in 2035 10% higher than the 2024 edition did, mainly for power in China and the United States, and natural gas 8 percent higher. UNEP records the same shift for U.S. emissions. In the same report, UNEP’s 2030 projection for U.S. emissions under current policies rose by 1 Gt of CO2-equivalent from the prior edition, largely because of policy reversals. The Global Carbon Budget projects a U.S. emissions rise in the latest year. U.S. fossil CO2 emissions are projected to rise 2.5 percent in 2025, reversing long-term decreases.

Which projections to trust

The scorecard suggests a simple ranking.

The physics deserves the most trust, conditional on the inputs. Temperature projections scored per unit of realized forcing have held up, and concentration projections have held up given emissions. That trust belongs to the assessed ranges, not raw model output. AR6 finds the CMIP6 models’ range of equilibrium climate sensitivity broader than the very likely range it assesses by combining multiple lines of evidence. The open question is sensitivity at the high end, which the acceleration debate keeps alive.

Aggregate energy demand and the fossil share of total energy deserve moderate trust at horizons of a decade or two. The IEA’s current-policies case got these right while missing everything inside electricity.

The composition of electricity supply deserves the least trust. Projections of solar and wind have missed by multiples for two decades, and the error has always run in the same direction. The cost trend that Way and coauthors fit has done better than the models; a reader should weight it more heavily than any scenario’s cost assumption.

Emissions projections deserve moderate trust at short horizons and little beyond a decade. Their error has run in both directions, and it has turned on growth in a single economy that no model forecast well.

For the SSPs specifically, the lesson is to treat SSP5-8.5 and SSP3-7.0 as stress tests, not baselines, for the next few decades. The plausible central paths now cluster around SSP2-4.5, which on our numbers sat close to the latest observed emissions.

The commentary view worth weighing is the dissent on China. If new Chinese power investment tilts back toward fossil plants as growth slows, the cyclical reading gains ground.²⁸ The one new climate-policy proposal in recent commentary is small relative to global emissions.²⁹

Methods and limits

What was pre-specified. The numbers analysis was planned before the data were examined. The growth-rate comparisons across periods, the cross-checks on CO2 growth and the re-basing of Hansen’s call are exploratory; the rest was pre-specified. Global fossil CO2 grew 1.1% a year over 1990-2000.

Data. Temperature is the mean of GISTEMP v4, Berkeley Earth and HadCRUT5, each re-based to a common pre-industrial baseline. Emissions are the Global Carbon Budget territorial fossil series, net of cement carbonation. Concentrations are NOAA’s global marine-surface means. Generation is Ember. Scenario paths come from the IPCC’s SRES files and the RCMIP harmonized RCP and SSP data. Past temperature projections come from the Hausfather digitization. WEO projections were read from each edition’s World annex tables.¹

Identity checks. The Kaya terms sum to the total change by construction, and the emitter contributions sum to the world shortfall. Our concentration for the previous year matches the published figure, and our fossil emissions match the Budget’s published total. The source put 2024 global mean CO2 at about 422.8 ppm; the frozen NOAA GML global annual mean is 422.79 ppm (exploratory check).

What the analysis cannot determine. Four limits matter most. First, the Kaya decomposition is an accounting split, not a causal one. Second, the trend-gap counterfactual assumes the early-century trends would have continued. For China that assumption is generous, so its share is better read in relative than absolute terms. Third, the wind-and-solar bounds assume the new generation displaced fossil generation one for one. If it partly met new demand, it displaced less. Fourth, the temperature scoring beyond the published windows is as issued only, because we did not assemble the forcing series a forcing-adjusted update would need.

What is missing. The central forcing-adjusted test is the Hausfather paper, which the analysis used through its published data and the AR6 chapter’s summary. The Peters et al. (2013) scoring of the early miss and Creutzig et al. (2017) on solar are available only in summary. We could not score the IS92 scenarios or the AR5 and AR6 near-term envelopes from machine-readable files, so the AR4 envelope stands in for the latter. We did not score the IEA’s stated-policies cases against outcomes, so the energy-mix scorecard is for reference and current-policies cases only. And we could not size explicit carbon pricing, so that part of the attribution rests on the literature.

What to monitor

• China’s emissions in the coming year. A year of flat or falling emissions with GDP growth near its recent pace would support the structural reading. A rebound with a growth recovery would support the cyclical one.
• The annual indicators update. The next Forster update will show whether the warming rate tracks the upper end of AR6’s range or exceeds it, and whether aerosol estimates move toward Hansen’s. PIIE’s Insider newsletter headlined that a super El Niño has arrived. A strong El Niño lifts temperatures and the CO2 growth rate for a year or two, whatever the trend.
• The IEA’s next World Energy Outlook. Whether the upward revisions to coal and gas projections continue, and whether solar additions again exceed the current-policies case. It arrives during an energy-price shock. CFR’s Benn Steil attributes the past year’s rise in U.S. and euro-area inflation to higher energy prices tied to the stalemate in the Strait of Hormuz.
• U.S. renewables deployment under the policy rollback. A modest slowdown would favor the cost explanation of the post-turn decline; a sharp one would favor the policy explanation.
• The land sink. A return of the airborne fraction toward its long-run average after the El Niño would confirm that the latest jump was weather. A persistent rise would be an unwelcome carbon-cycle surprise.
• Methane. Whether growth slows as the methane pledges take effect, or stays near the high scenarios, with the wetland feedback a growing share.

Appendix

This appendix sets out every series and transform the numbers analysis computed.

Temperature projections scored against the record

Fossil CO2 emissions against the scenarios

Global fossil CO2 emissions were 37.8 Gt of CO2 in 2024 (Global Carbon Budget 2025, net of cement carbonation). Each cell below is the scenario’s projection for that year relative to the outcome; positive means the scenario was above.

Emitters against their own trends

Kaya decomposition

Concentrations and the airborne fraction

The energy mix against the IEA World Energy Outlook

Bounds on the carbon-intensity channels

Horizon errors

The full mean-absolute-error table is the scorecard in the body; the electricity rows not shown there are gas generation at the three horizons: 10%, 7% and 14%.