IESET.
Hypotheses·energy·solar_lcoe_2010_2024_learning_curve_continuation

The photovoltaic (PV) learning curve — log cost of utility-scale solar modules and levelised cost of electricity (LCOE) declining linearly in log cumulative installed capacity at a learning rate of approximately 20-25% per doubling — continued through the 2020-2024 period despite (a) the COVID supply-chain shock 2020-2022, (b) the 2022 polysilicon + freight-rate spike, (c) the 2022-2023 inflation shock that reversed cost declines in many other capital-equipment classes, (d) US + EU trade defences against Chinese modules.

Specifically, global utility- scale solar module prices ($/W) and LCOE ($/MWh) by 2024 were lower than 2019 levels in real terms despite the macro inflation shock, and the implied 2010-2024 learning rate is statistically indistinguishable from the 2010-2019 learning rate. This is a sharp test against the alternative hypothesis that the PV learning curve flattened post-2020 due to commoditisation + supply-chain pressures.

INCONCLUSIVEengine/runs/solar_lcoe_2010_2024_learning_curve_continuation

INCONCLUSIVE_DATA_PENDING — no outcome variable loaded; missing: ['constructed: BloombergNEF / PV InfoLink polysilicon spot index, monthly. Manual-drop pending under data/manual/derived/.', 'constructed: ITRPV (International Technology Roadmap for Photovoltaic) annual industry-average mono-PERC + TOPCon + heterojunction efficiency. Manual-drop pending.']

confidence cueResult card produced; verdict unclassified.

policy briefCoverage too thin

In ordinary language

In plain terms, this asks whether log cumulative installed solar capacity gw is actually linked to better or worse log solar pv module price usd per w from 2010 to 2024.

plain answer

This test cannot make a firm call yet. no outcome variable loaded; missing: ['constructed: BloombergNEF / PV InfoLink polysilicon spot index, monthly.

why it matters

This matters because energy claims should change belief only when they survive a pre-declared empirical test.

how the test works

It compares 12 country or place units from 2010 to 2024, using a panel fe design.

what was measured
What changed
  • Log cumulative installed solar capacity gw
  • Post 2020 inflation shock dummy
What we checked
  • Log solar pv module price usd per w
  • Log solar pv lcoe usd per mwh
  • Log polysilicon spot price usd per kg
what this does not prove

A single test is not the whole truth. It narrows the claim under a specific sample, time period, and method. Strong policy conclusions need the pattern to survive nearby tests, alternative data, and serious objections.

verification

No evidence packet has been generated yet.

Results

engine/runs/solar_lcoe_2010_2024_learning_curve_continuation
1007550250201020172024CHNUSADEUINDJPNKORESP
illustrative sketch · run pending
No coefficients yet. When the model fires, this chart will show log_solar_pv_module_price_usd_per_w across 12 sampled countries over 2010–2024.
The shapes above are stylised — none of the lines are real data.
Placeholder for solar_lcoe_2010_2024_learning_curve_continuation. Published chart will be generated from engine/runs/solar_lcoe_2010_2024_learning_curve_continuation/chart_data.json.

Pre-registration

registration ordering unverified
first-spec commit 4c8ce8e · 2026-07-18T22:11:21Z
run generated · 2026-06-29T17:52:10Z
Run timestamp predates this path's first git-add commit (rebase, rename, or pre-git local run). Spec hash is still the path's first-add commit — not repository HEAD — but ordering is not a clean pre-registration proof.

The photovoltaic (PV) learning curve — log cost of utility-scale solar modules and levelised cost of electricity (LCOE) declining linearly in log cumulative installed capacity at a learning rate of approximately 20-25% per doubling — continued through the 2020-2024 period despite (a) the COVID supply-chain shock 2020-2022, (b) the 2022 polysilicon + freight-rate spike, (c) the 2022-2023 inflation shock that reversed cost declines in many other capital-equipment classes, (d) US + EU trade defences against Chinese modules. Specifically, global utility- scale solar module prices ($/W) and LCOE ($/MWh) by 2024 were lower than 2019 levels in real terms despite the macro inflation shock, and the implied 2010-2024 learning rate is statistically indistinguishable from the 2010-2019 learning rate. This is a sharp test against the alternative hypothesis that the PV learning curve flattened post-2020 due to commoditisation + supply-chain pressures.

Falsification criterion — what would disprove this

set before the run · honoured after

This hypothesis is considered falsified if:

Not supported if (a) β_full < 0.7 × β_pre (learning rate flattened by more than 30%), OR (b) Chow-test for break in 2020-2022 rejects no-break at p<0.10 with break that is a slowdown, OR (c) module $/W in 2024 exceeds 2019 real $/W (no progress despite cumulative capacity tripling), OR (d) the BOS-cost component shows zero or positive learning rate (BOS-cost learning has historically been weaker; if it has flatlined entirely, LCOE story is more fragile).

formal test & threshold
test:      solar_learning_curve_2010_2024
threshold: β_full ≥ 0.85 × β_pre (≤15% flattening tolerated) AND No structural break at p<0.10 AND 2024 module $/W < 2019 module $/W in real 2023 USD AND BOS-cost learning rate > 0 over 2010-2024.

Method

Template
panel_fe
Fixed effects
Clustering
year
Sample
12 countries · 2010 – 2024
Evidence type
associational

Primary specification: log(cost) = α + β × log(cumulative capacity) + ε, estimated separately for two windows: 2010-2019 (pre-shock) and 2010- 2024 (full sample). Test: H0 β_full = β_pre, against H1 the coefficient has flattened (|β_full| < |β_pre|). Secondary: structural-break test (Chow / Quandt-Andrews) for break- point in log-cost on log-cumulative-capacity series 2020-2022. If no break detected, learning curve continued through inflation shock. Tertiary: residual analysis — plot residuals from 2010-2019 fitted learning curve against 2020-2024 actuals. Inflation-shock spike 2022 visible as positive residual; subsequent reversion below fitted line by 2024 indicates the trajectory caught up. Decomposition: separate module $/W (technology-cost) from BOS (balance-of-system) $/W (installation cost). The learning rate is historically faster on module than BOS; if the 2020-2024 module- learning continued but BOS flattened, the LCOE result is mixed. Known limitations: (1) IRENA cost data is annual; finer-resolution test using BloombergNEF monthly module spot would be sharper. (2) Trade-defended markets (US post-2018 Section 201 tariffs, EU 2013- 2018 anti-dumping) had higher domestic prices than world spot; the global learning curve mixes these regimes. (3) Capacity factor improvements (bifacial modules, tracker uptake) are part of LCOE decline that is not module-cost-curve. (4) Currency effects: USD-denominated cost vs CNY-denominated production cost; some 2022-2024 USD price stability reflects CNY weakness rather than CNY-cost decline.

Data

VariableSourceTransform
log_solar_pv_module_price_usd_per_w
outcome
irena:solar_pv_coststier 2
log
log_solar_pv_lcoe_usd_per_mwh
outcome
irena:lcoe_solar_pvtier 2
log
log_polysilicon_spot_price_usd_per_kg
outcome
constructed:BloombergNEF / PV InfoLink polysilicon spot index, monthly. Manual-drop pending under data/manual/derived/.tier 5
log
log_solar_module_efficiency_pct
outcome
constructed:ITRPV (International Technology Roadmap for Photovoltaic) annual industry-average mono-PERC + TOPCon + heterojunction eftier 5
log
log_cumulative_installed_solar_capacity_gw
treatment
irena:capacitytier 2
log
post_2020_inflation_shock_dummy
treatment
constructed:indicator = 1 from 2020-Q1 onwards (COVID supply-chain shock + subsequent commodity + freight inflation).tier 5
indicator
log_brent_oil
control
imf_pcps:POILBREtier 1
log
log_industrial_metals_index
control
imf_pcps:PMETAtier 1
log
log_us_cpi
control
fred:CPIAUCSLtier 1
log

● ready  · ● pending  · ● reconstruct-needed

Detailed result card

Result card — solar_lcoe_2010_2024_learning_curve_continuation

Verdict: INCONCLUSIVE_DATA_PENDING — no outcome variable loaded; missing: ['constructed: BloombergNEF / PV InfoLink polysilicon spot index, monthly. Manual-drop pending under data/manual/derived/.', 'constructed: ITRPV (International Technology Roadmap for Photovoltaic) annual industry-average mono-PERC + TOPCon + heterojunction efficiency. Manual-drop pending.']

Pre-registration

  • Claim: The photovoltaic (PV) learning curve — log cost of utility-scale solar modules and levelised cost of electricity (LCOE) declining linearly in log cumulative installed capacity at a learning rate of approximately 20-25% per doubling — continued through the 2020-2024 period despite (a) the COVID supply-chain shock 2020-2022, (b) the 2022 polysilicon + freight-rate spike, (c) the 2022-2023 inflation shock that reversed cost declines in many other capital-equipment classes, (d) US + EU trade defences against Chinese modules. Specifically, global utility- scale solar module prices ($/W) and LCOE ($/MWh) by 2024 were lower than 2019 levels in real terms despite the macro inflation shock, and the implied 2010-2024 learning rate is statistically indistinguishable from the 2010-2019 learning rate. This is a sharp test against the alternative hypothesis that the PV learning curve flattened post-2020 due to commoditisation + supply-chain pressures.
  • Falsification rule: Not supported if (a) β_full < 0.7 × β_pre (learning rate flattened by more than 30%), OR (b) Chow-test for break in 2020-2022 rejects no-break at p<0.10 with break that is a slowdown, OR (c) module $/W in 2024 exceeds 2019 real $/W (no progress despite cumulative capacity tripling), OR (d) the BOS-cost component shows zero or positive learning rate (BOS-cost learning has historically been weaker; if it has flatlined entirely, LCOE story is more fragile).
  • Falsification test: solar_learning_curve_2010_2024

Estimate

  • Error: no outcome variable loaded; missing: ['constructed: BloombergNEF / PV InfoLink polysilicon spot index, monthly. Manual-drop pending under data/manual/derived/.', 'constructed: ITRPV (International Technology Roadmap for Photovoltaic) annual industry-average mono-PERC + TOPCon + heterojunction efficiency. Manual-drop pending.']

Variables resolved

  • irena:solar_pv_costs → log_solar_pv_module_price_usd_per_w (outcome, publisher=irena, n=15)
  • irena:lcoe_solar_pv → log_solar_pv_lcoe_usd_per_mwh (outcome, publisher=irena, n=15)
  • irena:capacity → log_cumulative_installed_solar_capacity_gw (treatment, publisher=irena, n=5848)
  • constructed: indicator = 1 from 2020-Q1 onwards (COVID supply-chain shock + subsequent commodity + freight inflation). → post_2020_inflation_shock_dummy (treatment, publisher=constructed, n=180)
  • imf_pcps:POILBRE → log_brent_oil (controls, publisher=imf_pcps, n=444)
  • imf_pcps:PMETA → log_industrial_metals_index (controls, publisher=imf_pcps, n=444)
  • fred:CPIAUCSL → log_us_cpi (controls, publisher=fred, n=960)

Variables missing data

  • constructed: BloombergNEF / PV InfoLink polysilicon spot index, monthly. Manual-drop pending under data/manual/derived/. (outcome, name=log_polysilicon_spot_price_usd_per_kg) — vintage not on disk
  • constructed: ITRPV (International Technology Roadmap for Photovoltaic) annual industry-average mono-PERC + TOPCon + heterojunction efficiency. Manual-drop pending. (outcome, name=log_solar_module_efficiency_pct) — vintage not on disk

Generated by scripts/run_panel_fe.py at 2026-06-29T17:52:10+00:00

Strongest opposing argument

Every hypothesis ships with its charitable opposing argument. The framework earns credibility by handling objections at their strongest, not weakest.

Notes

Data readiness: - IRENA cost + capacity panels: ready - IMF PCPS (oil, metals): ready - FRED CPI: ready - BloombergNEF polysilicon + module spot: manual-drop pending under data/manual/derived/ - ITRPV efficiency: manual-drop pending Run when BNEF polysilicon manual-drop is populated; v1 can run on IRENA alone for the headline learning-rate test.

Authored framework. Read the transparency note.