Forecast Lab • v0
Claims that can resolve
Foresight reports orient a conversation, and the good ones do it well. But the genre is built not to resolve: trends have no dates, claims have no criteria, and no report is ever scored against what happened. This ledger is the opposite object. Every entry is a claim with a criterion and a resolution date fixed at logging time; when the date comes, the outcome is recorded publicly, misses exactly like hits. Some entries track other institutions' own timestamped claims (the scorekeeper role); a few stake this site's own reading (clearly separated); one is a retrospective worked example that is never counted.
The rules
- The logged date is when the entry entered the ledger; the site's git history is the audit trail.
- Criteria are fixed at logging and never edited. A criterion that turns out unjudgeable resolves as "unresolvable", and that failure is itself recorded.
- Backtests (claims logged after their outcome) are marked and never counted toward any tally.
- The ledger grows only with claims worth staking. No volume targets, no reweighting, no averages until the sample is honest.
Open claims (18)
Ledger updated 2026-07-16.
EUDR obligations apply on 30 December 2026 with no further postponement.
● OpenCOM(2026) 191 final (May 2026) states the Commission will not propose further amendments to the basic legal text and confirms application at the end of 2026.
Criterion No further postponement is enacted and large/medium-operator obligations are in application on 30 December 2026.
Indonesia, Malaysia and Brazil remain standard risk in the first EUDR country-benchmarking review.
● OpenThe first classification (22 May 2025) put all three at standard risk; the first review is due during 2026.
Criterion The published review keeps all three countries in the standard-risk class (or no review is published in 2026, resolving unresolvable).
GE Vernova's gas-turbine reservations are effectively sold out through 2030 by the end of 2026.
● OpenDecember 2025 investor update: ~80 GW backlog into 2029, slot reservations expected sold out through 2030 by end-2026.
Criterion GE Vernova's Q4/FY2026 reporting or guidance states heavy-duty gas-turbine slots are sold out (or fully reserved) through 2030.
Germany's hydrogen Kernnetz has its first ~525 km of segments in operation by the end of 2026.
● OpenIEA GHR 2025 reported the first 525 km (mostly repurposed gas lines) due in 2025; not yet confirmed commissioned, so the claim is tracked with a year's grace.
Criterion Official confirmation (FNB Gas, BNetzA or operator) that at least ~500 km of Kernnetz segments are in hydrogen operation.
Keliber produces its first battery-grade lithium hydroxide at Kokkola during 2026.
● OpenMining started February 2026; the refinery is in staged commissioning per company reporting cited on the Finnish mining page.
Criterion Sibanye-Stillwater reporting confirms first battery-grade lithium hydroxide produced at Kokkola with a 2026 date.
If Terrafame's reported 2026 sulphate discharge swings more than 35 percent from 2025, the independently measured above-background sulphate flux at the Nuasjarvi outflow moves the same direction by at least 20 percent.
● OpenThe pre-registered test of the first dependency edge (edge:terrafame-reported-so4--nuasjarvi-measured-flux, src/data/registry/edges.json). The 2016-2025 retrospective on the mine-water page formed the hypothesis: the lake corroborated the one large reported swing (the 2018 closed-cycle collapse) but not year-to-year variation, so the registered test is directional and fires only on a qualifying swing. The measured side (SYKE VESLA and Hydrology, open data) is computable in early 2027 and will be locked before Terrafame's report publishes; that ordering is the independence of the test.
Criterion Qualifying swing: Sustainability Report 2026 puts sulphate in discharges outside 6,557-13,619 t (35 percent around 2025's 10,088 t). If it qualifies: measured 2026 above-background outflow flux (balance.json method) at or beyond 28,359 t on an upswing, or at or below 18,906 t on a downswing (20 percent around the 2023-2025 mean of 23,632 t), scores a hit; opposite movement or a missed threshold scores a miss and refutes the edge for annual resolution. No qualifying swing, or fewer than 5 SO4-sampled months in the measured year, resolves unresolvable-by-design and the test re-registers for 2027.
Terrafame's 2026 sulphate load in discharged water stays under the 15,780 t/yr permit quota.
● OpenThe mine-water worked example: reported loads of 12,424 t (2024) and 10,088 t (2025), a demonstrated closed-cycle capability, and treatment holding the metals back. The stake is that the reported-versus-real structure we verified continues to hold.
Criterion Terrafame's Sustainability Report 2026 reports sulphate in discharges below 15,780 t.
European hydrogen pipeline kilometres under construction stay below 500 km at GEM's next Europe hydrogen update.
● OpenThe hydrogen-pipelines page's demand-side read: 50,165 km proposed against 30 km under construction, under 6% of announced kilometres at FID globally, and a widening cost gap. The stake is that construction, not announcements, is the binding series.
Criterion GEM's next Europe Gas Tracker hydrogen edition (expected early 2027) reports under 500 km under construction in Europe.
At least one direct PPA is executed under Thailand's 2,000 MW pilot by the end of 2026.
● OpenThe power-markets layer: pilot approved June 2024, draft direct-PPA and third-party-access regulations issued October 2025, scoped first to BOI-promoted data centres, where demand pressure is real. Lower confidence: Thai regulatory timelines slip.
Criterion ERC, EGAT or a counterparty publicly confirms a signed direct PPA under the pilot dated 2026.
The UAE commissions the second, ~1.5 million b/d Jebel Dhanna to Fujairah crude line by the end of 2027.
● OpenThe EIA's March 2026 chokepoints brief records the planned line, which would roughly double the UAE's Hormuz-bypass capacity.
Criterion Official commissioning or start-of-operations announcement for the second Fujairah crude line dated 2027 or earlier.
Kazakh oil transit to Germany's Schwedt refinery via Druzhba does not resume during 2026.
● OpenRussia announced the halt of Kazakh (KEBCO) transit from 1 May 2026, citing technical capacities. The entry scores whether that stated halt holds through the year.
Criterion Public reporting confirms no resumed KEBCO deliveries via Druzhba to the PCK Schwedt refinery in calendar 2026.
The CPC pipeline suffers at least one further capacity-reducing disruption (roughly 20 percent or more, lasting three days or longer) before 30 June 2027.
● OpenThe oil-pipelines page's leverage read: a corridor carrying the large majority of Kazakhstan's exports across Russian territory to one terminal was disrupted three times in 2025 alone (drone strike, regulatory mooring suspension, terminal strike). The stake is that the structure, not any single event, drives recurrence.
Criterion Operator statements or wire reporting document at least one qualifying disruption between logging and the resolution date.
Electric cars reach 53 percent of world new-car sales in 2035 (IEA STEPS).
● OpenThe GEVO 2026 Stated Policies Scenario projects a 53% world EV sales share for cars in 2035 (46% under Current Policies). Logged from the vintage archived on the EV observatory, because the IEA's live dataset is overwritten each edition; this entry freezes what was actually projected in May 2026.
Criterion The realized 2035 world EV share of new car sales (BEV+PHEV, IEA or equivalent successor series) lands within 5 percentage points of 53. Above 58 scores the projection an underestimate, below 48 an overestimate, in the solar-board sense.
The global carbon credit market reaches USD 4,734 billion in value by 2030.
● OpenA vendor projection of global carbon-market value (about 99 percent of it compliance allowances in GVR's own segmentation), at a 38.7 percent CAGR. Logged from the archived June-2023 wording because the live press-release URL has already been silently overwritten with a new 'USD 6,130 billion by 2033' vintage, which is itself part of what this entry records. For scale: global traded value was roughly EUR 881 billion in 2023 and just above EUR 800 billion in 2024 (LSEG), so the claim needs the market to roughly quintuple in six years.
Criterion LSEG's Carbon Market Year in Review for 2030 (published early 2031) reports global carbon-market traded value of at least USD 4.7 trillion or the euro equivalent. Government carbon-pricing revenue series (World Bank) are not the measuring stick.
CORSIA-eligible credit demand structurally outstrips supply by 2030.
● OpenAbatable's scenario analysis put cumulative Phase 2 demand at 14 times supply if no new eligible projects emerge (7 times under its optimistic-demand case), with the crossover between 2029 and 2030. The headline multiple is a conditional counterfactual, so the entry scores the direction, not the number. The direction already shows: as of September 2025 IATA counted 15.8 million eligible units available against 146 to 236 million needed for Phase 1 alone.
Criterion ICAO registry and IATA data as of end-2030 show cumulative CORSIA-eligible unit supply (issued, with host-country letters of authorization) below cumulative offsetting requirements through 2030. Eligible supply meeting demand scores a miss.
Finland's BEV share of new-car sales dips in 2024 from its 2023 level.
● HitWorked example of resolution mechanics, logged after the fact: an A1AYN deck written in spring 2024 flagged the drop in Finnish BEV sales as a near-term cloud over EV adoption, when the first quarter stood at 25.8 percent against 33.8 for full-year 2023. Scored on the register's own series via the EV observatory. Not counted in any tally.
Criterion Traficom-registered BEV share of new passenger cars in full-year 2024 comes in below the 2023 share.
Outcome Full-year 2024 BEV share was 29.5 percent against 33.8 in 2023 (Traficom). The dip was real and temporary: 2025 recovered to 37.2 percent, above the 2023 level, which is the part the caution could not see.
Critical-mineral extraction and refining consolidate progressively around lowest-cost producers over 2024 to 2026.
● HitLogged retrospectively as a worked example and verified mid-2026 against company disclosures and the IEA's concentration figures. Not counted in any tally.
Criterion Documented curtailments or exits among higher-cost producers, with the supply share of the lowest-cost producers and refiners rising, per IEA concentration figures and company disclosures by mid-2026.
Outcome Largely as read: Australia's nickel industry all but shut (BHP's Nickel West suspended October 2024, Glencore's Koniambo to care and maintenance February 2024) while Indonesia passed roughly 60 percent of mine supply; the 80-plus percent lithium crash forced curtailments and consolidation, Rio Tinto closing the 6.7 billion dollar Arcadium acquisition in March 2025; the IEA puts the average top-three refining share at 86 percent in 2024, up from 82 in 2020. Two caveats recorded with the hit: several 2025 concentration events were state-directed rather than cost-driven (DRC cobalt quotas, CATL's licence halt), and the late-2025 lithium price recovery began pulling higher-cost supply back.
A new-build export hydrogen pipeline without committed buyers does not survive to FID.
● HitWorked example of resolution mechanics only, logged after the fact: Equinor and RWE cancelled the Norway-Germany hydrogen pipeline in September 2024, explicitly for lack of viable demand, matching the demand-side read now on the hydrogen page. Not counted in any tally.
Criterion Project cancelled or indefinitely shelved before FID.
Outcome Cancelled September 2024; Equinor cited absent long-term buyer commitments.
The Talvivaara near-field lakes' sulphate follows the mine's operational arc: near-zero before mining, hundreds to thousands of mg/l during the 2010-2014 crisis loading, then recovery after the November 2015 rerouting of discharge to the Nuasjarvi pipe.
● HitThe upstream-lakes edge (edge:talvivaara-discharge--upstream-lakes-sulphate). SYKE VESLA open sulphate data at Kalliojarvi and Kivijarvi, the southern receptors, read as a natural experiment. Not a forward test (the events are historical), so logged as a backtest and never counted in any tally. Honest caveat: the post-2015 recovery coincides with Terrafame's treatment upgrade, so it reflects the whole operational change, not the pipe diversion alone.
Criterion Kalliojarvi and Kivijarvi sulphate rises from a few mg/l pre-mine to hundreds or thousands during 2010-2014, then falls back toward low hundreds or below within about three years of the 2015 rerouting.
Outcome As read, and cleanly. Kalliojarvi: 2-3 mg/l (2008-2009), peak 3,650 (2012), down to 335 (2016) and 97 (2017). Kivijarvi: 3 mg/l (2008), 780-1,300 (2010-2014), down to 270 (2016) and 49 (2018). Salminen went meromictic and stayed saline far longer (8,000-12,500 mg/l through 2020), recovering only after its deep layer was remediated (38 mg/l by 2023), which is itself the signal that a trapped saline layer does not flush on its own.
Brent stays below $100 through 2026-10-31 unless verified Hormuz throughput (PortWatch, 7-day average) falls below 5 daily transit calls for 14 or more consecutive days.
● OpenThe price-transmission page's buffer-absorption thesis: Brent peaked at $138 (7 April 2026), currently at $70. The IEA coordinated release of 400 Mbbl (11 March), approximately 5 mb/d of OPEC+ spare capacity, demand destruction (-420 kb/d y/y per IEA OMR), and bypass pipelines (~9 mb/d) are anchoring crude. Current Hormuz transit is approximately 10 daily calls (vs ~72 pre-crisis). The claim stakes that these buffers hold unless transit collapses further toward zero.
Criterion Brent settles below $100/bbl on every trading day through 2026-10-31, OR PortWatch Hormuz 7-day average transit calls fall below 5 for 14 or more consecutive days (at which point the condition is met and the claim is not falsified regardless of price). Falsifier: Brent settles above $100 for 5 consecutive trading days without the transit condition being met.
US Atlantic Coast diesel crack (EIA NY Harbor ULSD minus Brent, $/bbl) stays above $35/bbl through 2026-10-31 or until JODI shows Asia+ME refinery intake within 1.5 mb/d of the January 2026 level.
● OpenThe price-transmission page's downstream-migration thesis: diesel crack went from approximately $28/bbl (January 2026 average) to $61-69/bbl (July 2026), driven by 4-6 mb/d of Asia/ME refinery run cuts behind the chokepoint. The crack will not normalise until refinery intake recovers. The claim stakes the thesis that the binding constraint migrated from the strait to the refinery slate, and the signal lives in cracks, not in crude.
Criterion EIA-computed diesel crack (NY Harbor ULSD in $/bbl minus Brent) stays above $35/bbl on every trading day through 2026-10-31, OR JODI monthly data shows Asia+ME refinery intake within 1.5 mb/d of the January 2026 level (at which point the refinery-recovery condition is met and the claim resolves regardless of crack level). Falsifier: cracks normalise below $35/bbl for 10 consecutive trading days while JODI Asia+ME intake remains more than 1.5 mb/d below January 2026.
The EU and India conclude a free trade agreement by the end of 2026.
● HitLogged as a backtest because the outcome was already known at logging time. The India market-entry plan (drafted 16 Jul 2026) listed 'EU-India FTA concluded by a dated horizon' as a forecast claim. During execution, web research revealed the FTA had been concluded on 27 January 2026, resolving the claim before it was even registered. Honest accounting: we would have gotten this right only because it happened before we asked the question.
Criterion EU and India announce the conclusion of FTA negotiations.
Outcome Concluded 27 January 2026. The EU's largest FTA by population coverage (~2 billion people, ~25% of global GDP). Covers goods, services, and regulatory issues; excludes investment chapter and comprehensive GI agreement. Pending ratification by both sides.
India's all-India peak power deficit stays below 3% in FY 2026-27.
● OpenCEA Load Generation Balance Report 2025-26 projected a 1.2% peak deficit (3,271 MW shortfall against 269,277 MW demand). The trend is improving: renewable capacity additions are accelerating, dedicated freight corridor electrification reduces diesel traction load, and Grid-India's interstate transfer capability has expanded. The binding constraint is not generation capacity but intra-day ramping and sub-regional transmission, which is why we use the national aggregate, not a state-level claim.
Criterion CEA's Monthly Executive Summary or the Load Generation Balance Report for FY 2026-27 reports all-India peak deficit below 3% for every reported month. Falsifier: any month exceeds 3%.
The board: EIA's oil forecasts, scored
Computed 2026-07-02 from 29 STEO release vintages (2018 to 2025). ● measured
The ledger's quantitative sibling: the US EIA's Short-Term Energy Outlook, scored against what happened and against a bench of naive rivals, on ~500 forecast-versus-actual pairs per series at horizons up to 18 months. Every forecaster, the baselines included, sees only the history known at that release. The honest headline runs both ways: on oil prices STEO beats a random walk by single-digit percent, prices are close to unforecastable; on US crude production its skill is large and real. Skill is measured against simply assuming no change. The small chart in each card shows where that edge lives: the gap between the green line (STEO's average miss) and the dashed grey bar (no-change) is the forecaster's added information, horizon by horizon; where the lines meet, forecasting has run out of road.
How to read this
One scored pair works like this. In January 2019 the EIA published a forecast of the July 2019 WTI price, a 6-month horizon. July came, the real price was recorded, and the difference between the two is that pair's error. Repeat for every forecast month in every archived release (each old release is a vintage), and about 500 such pairs accumulate per series. Two numbers summarise them:
skill = 1 − MAE / MAEno-change (share by which a forecaster beats "assume nothing changes")
"No-change" simply repeats the last known value at every horizon; it is the bar every real forecaster must clear, because anyone can forecast that tomorrow looks like today. The other rivals (drift, moving average, Holt smoothing, a Monte Carlo simulation) are equally naive recipes run on the same information, which is the point: a forecaster only earns credit for what beats knowing nothing.
What makes one forecaster better than another. Better means a smaller average miss on identical fixtures, nothing else. The sign of the skill number is the sharp edge: positive skill means the forecaster added information beyond the last known value; negative skill means all its analysis produced forecasts worse than doing nothing at all, you would literally have lost accuracy by listening to it. That is why most of the bench sits below zero on the price boards: oil prices are close to a random walk, and effort without information subtracts value. Magnitude matters too: +6 percent on a near-unforecastable series is a modest, real edge; +36 percent on production is a decisive one. Skill compares forecasters within one series, never across series.
Accuracy is also not the whole story. Two forecasters can tie on MAE while one is systematically biased, always a little high or always a little low (bias = the signed average error). A tight but biased forecaster is the more useful one, because a known lean can be corrected by subtracting it; an unbiased but wild one cannot be fixed. The board tracks both, and a persistent bias sustained over years is the sharpest thing a scorekeeper can catch, it means the forecaster's model of the world leans, not just its luck.
So a row like "EIA STEO 13.36 (+6%)" reads: across ~500 pairs, STEO's average miss on the WTI price was $13.36 per barrel, a 6 percent smaller miss than assuming no change. Green means the forecaster beat that bar; "common sample" means every forecaster is scored on exactly the same pairs, so nobody plays easier fixtures. No forecaster, STEO included, ever sees data published after its release date.
WTI crude price
MAE in $/bbl · n = 501 pairs · common sample
Leader's bias: -1.289 $/bbl (essentially unbiased)
— STEO vs - - no-change, average miss by months ahead. MC 80% band covered 90% of actuals (bands run wide: honest, if cautious).
Brent crude price
MAE in $/bbl · n = 501 pairs · common sample
Leader's bias: -1.351 $/bbl (essentially unbiased)
— STEO vs - - no-change, average miss by months ahead. MC 80% band covered 88% of actuals (bands run wide: honest, if cautious).
US crude production
MAE in million bbl/day · n = 499 pairs · common sample
Leader's bias: +0.005 million bbl/day (essentially unbiased)
— STEO vs - - no-change, average miss by months ahead. MC 80% band covered 78% of actuals (well calibrated).
Prices (WTI, Brent) are effectively revision-immune; US crude production is scored against the current revised EIA series, not the vintage known at the time, so its errors mix forecast error with data revisions. The full method, the per-horizon breakdown and the interval-calibration check live in the Forecast Lab workshop; finished boards are copied here as frozen snapshots, never computed in the build.
The counter-exhibit: the WEO's solar projections, scored
Computed 2026-07-03; forecasts from the CC-BY compilation of Lopez, Pourjamal & Breyer (2025); realized capacity to 2024. ● measured
The board above showed a well-calibrated forecaster: EIA's short-term outlook, essentially unbiased. Here is the same lesson's other half, from the IEA's long-term World Energy Outlook. Each vintage's central scenario projected world solar capacity at five-year milestones; 10 of those projections have now resolved, and every single one undershot reality, by 56 percent on average, and by about 49 percent even at the five-year horizon, where policy in force at release largely determines the build-out. Ten misses is a small sample for measuring error size; ten misses all in the same direction is a sign test, and that is the point. This is bias, a model of the world that leans, not bad luck.
Signed error of every resolved projection (vintage → target year)
Worst: WEO 2010 projected 110 GW for 2020; reality delivered 720 GW. The 2017–2022 vintages resolve at their 2025 and 2030 milestones; the board grows as they do.
IEA states its scenarios are conditional on policy, not forecasts. The central scenario is scored because it is the one read as the outlook by markets and governments; short-horizon (about five-year) errors are reported separately because policy in force at release largely determines a five-year build-out.
The dependency edges
Where a ledger entry stakes this site's own reading, the hypothesis behind it lives in a public registry of dependency edges: what moves what, through which mechanism, with which sign and lag, and with the falsifier written before the data can resolve it. An edge starts conjectured, earns corroborated only by surviving an out-of-sample test against named null models, and is marked refuted when it fails; refuted edges stay on the record. The registry currently holds the mine-water edges, including the two where the honest verdict is that no severe test passes yet, which is itself a finding worth publishing.
Terrafame's reported sulphate discharge is verifiable in the SYKE-measured Nuasjarvi outflow flux
forward-test● Under testDriver Reported annual sulphate discharge, Terrafame primary disclosures ● measured
↓ sign + · lag 0-1 years, lake hydraulic residence
Responder Above-background sulphate flux at the lake outflow, SYKE VESLA concentration times SYKE Hydrology discharge, Koivukoski Paikka 1303 ● measured
Mechanism Bioheap leaching of sulphide ore mobilises sulphate; water treatment precipitates the metals while the sulphate passes through to the discharge pipe, and what enters the lake must transit the outflow within the hydraulic residence time. Discharged load is therefore a component of the measured outflow flux, riding on catchment background (pre-mine baseline 7.4 mg/l).
Evidence, magnitude and the registered test ▾
Magnitude Reported 3,400-13,600 t/yr against measured above-background flux 15,900-65,800 t/yr over 2016-2025; background and hydrology dominate, so only reported swings larger than roughly 35 percent are expected to be resolvable. The measured-to-reported ratio ranged 1.7 to 6.3, which is why the test is directional, not magnitude-based.
Evidence window 2016-2025 retrospective. This window formed the hypothesis and is not reusable as its test.
Reading The lake corroborates the one large reported swing: the 2018 closed-cycle collapse to 3,434 t shows as the 2018-2019 trough in measured flux (18,289 and 17,941 t against neighbouring years at 39,000-66,000). It does not resolve year-to-year variation: 2025 illustrates the limit, a dry year with 4 sampled months where measured flux (15,882 t) fell below the 2018 trough while reported discharge was normal (10,088 t).
Prediction If Terrafame's reported 2026 sulphate discharge differs from the reported 2025 figure (10,088 t) by more than 35 percent in either direction, the SYKE-measured above-background outflow flux for 2026, computed by the balance.json method, differs from the 2023-2025 measured mean (23,632 t) by at least 20 percent in the same direction.
Decision rule Qualifying swing: reported 2026 outside 6,557-13,619 t. On a qualifying swing, measured 2026 above-background flux must land at or beyond 28,359 t (upswing) or at or below 18,906 t (downswing). Fewer than 5 SO4-sampled months in the measured year resolves unresolvable; 2025 itself had 4.
Falsifier On a qualifying reported swing with adequate sampling, the measured flux moves in the opposite direction or fails the 20 percent same-direction threshold. The edge is then refuted for annual-resolution verification at this site.
Registered 2026-07-07 · resolves by 2027-05-31 · ledger entry terrafame-lake-covariation-2026
Terrafame's reported discharge covaries quarter-to-quarter with discharge-bay conductivity
mechanism● ConjecturedDriver Reported quarterly sulphate discharge, Terrafame monthly water-emissions reports ● measured
↓ sign + · lag 0-1 quarters
Responder Discharge-bay conductivity (Jormaslahti / Nuasjarvi bay stations, bottom water where the dense plume pools), SYKE VESLA ● measured
Mechanism The pipe discharges into Jormaslahti bay; the treated water is dense and saline, so it sinks and pools on the bay bed. The bay is where the plume is strongest (measured sulphate enrichment about 22x the open lake), so unlike the catchment-diluted outflow the bay is not swamped by background. A quarter's discharge should therefore show in the bay before it dilutes lake-wide. This was the proposed positive-test responder, chosen because the outflow edge above is a deliberate negative control.
Evidence, magnitude and the path to a testable form ▾
Magnitude Bay sulphate runs ~22x the open lake (measured), so the spatial signal is large; but the temporal covariation with the reported load is not.
Evidence window 2016-2025, reported quarterly pipe load (discharge_monthly.json) against SYKE VESLA bay conductivity and sulphate.
Reading Confirms that no receptor at this lake, outflow or bay, supports a naive quarterly positive test with the reported load. The verification gap is deeper than the outflow alone: the open-data read localises the plume in space but cannot resolve the discharge in time, because the driver is too stable and every receptor is physically confounded.
Severity check Checked against data before building, per the H1->H1' precedent, and the naive quarterly covariation FAILS at every candidate responder: reported pipe load vs bay bottom conductivity r=0.02, vs bay surface sulphate r=0.01, vs bay-minus-open-lake reference gradient (the BACI control) r=0.04 on levels and -0.18 on quarter-to-quarter changes, and even the annual bay-minus-reference gradient is only r=0.15. Two reasons, both physical: (1) the reported driver barely varies within the pipe era (10,000-13,600 t/yr except the 2018-2019 closed-cycle dip), so there is little to correlate against, the same low-severity problem that sank the original production-nowcast H1; (2) the bay signal is governed by the stratification cycle (accumulate under ice, flush at turnover) and plume residence, not by that quarter's discharge, and grab sampling is sparse (~2 samples/quarter). The spatial signal (22x enrichment) is strong; the temporal covariation is not.
Path to a testable form Two testable forms the data could support, neither a naive covariation: (a) EVENT-based directional, like the outflow edge, does a future large sustained discharge cut (a closed-cycle-like event) drop bay bottom conductivity within 1-2 quarters; the bay should be more sensitive than the outflow, but only a real swing can test it, and the driver has not swung since 2019. (b) The UPSTREAM receptor lakes (Kivijarvi, Salminen, Kalliojarvi) where Terrafame production actually swung hard during the 2012-2015 crisis and bankruptcy, a different edge with a driver that genuinely varies. Registered conjectured, not under-test, because no severe test currently passes.
Talvivaara's crisis-era discharge, and its 2015 rerouting, is written large in the upstream receptor lakes
natural-experiment● ConjecturedDriver Talvivaara/Terrafame effluent routing and loading, 2008-2018 (the 2012 gypsum-pond leak, crisis discharges, and the Nov 2015 diversion to the Nuasjarvi pipe) ● measured
↓ sign + · lag weeks to months (direct near-field receptors)
Responder Upstream receptor-lake sulphate, SYKE VESLA (Salminen, Kalliojarvi, Kivijarvi, south and near-field of the mine) ● measured
Mechanism Before the Nuasjarvi pipe (Nov 2015) the mine's sulphate-metal load, plus the Nov 2012 gypsum-pond leak, discharged into the small southern and near-field lakes. Sulphide leaching loads sulphate; in small lakes the dense saline water sinks and, in Salminen, traps as a meromictic deep layer that does not flush. The 2015 pipe rerouted effluent north, so the southern lakes should recover while Salminen's deep layer stays saline.
Evidence, magnitude and the path to a testable form ▾
Magnitude Enormous and unambiguous: Salminen SO4 2 mg/l (2009) to 8,000-11,000 (2011-2015); Kalliojarvi to a 3,700 peak (2012) then recovery to ~98 (2017); Kivijarvi to 1,300-1,700 (2014) then ~45-50 (2018). 100-4,000x background, versus the ~1.5x lake-wide and 22x discharge-bay at Nuasjarvi.
Evidence window 2009-2018 SYKE VESLA SO4 at the near-field lakes (pulled 2026-07-08).
Scoping verdict A spectacular natural experiment, but its epistemic character is retrospective, not forward. Three findings from the data: (1) the responder is a step-change plus accumulation, not a proportional covariation, Salminen went meromictic and stays saline regardless of later discharge, so 'production -> sulphate' year-by-year does not hold; (2) the driver is confounded, the dominant loading events (the 2012 gypsum-pond leak, crisis discharges) were accidents decoupled from production RATE, and production fell while pollution rose; (3) the cleanest signal is the 2015 INTERVENTION, the pipe rerouting north, followed by sharp southern-lake recovery (Kalliojarvi 2,600 -> 520 -> 98; Kivijarvi ~1,500 -> ~250 -> ~48), though this is confounded with Terrafame's concurrent treatment upgrade.
Role Complementary to the Nuasjarvi forward edge, not a replacement. Upstream = the large-signal RETROSPECTIVE proof that open water data reads a mine's operational change unambiguously WHEN the driver swings hard (the proof-of-concept for the paper). Nuasjarvi outflow = the honest FORWARD pre-registered test on a now-stable operation. The method's power and its limits, on one mine.
Path to a testable form No severe FORWARD test exists here: the crisis is over, the lakes have recovered or gone meromictic, and the driver is now stable, so a forward prediction would be low-severity persistence. The resolved retrospective sub-finding (2015 rerouting -> southern-lake recovery) is registrable as a ledger BACKTEST (kind: backtest, never counted in any tally), which is the honest home for an in-sample worked example. Registered conjectured here because the forward version is weak.
Ren-Gas Tampere FID requires environmental permit (resolved)
statutory● ConjecturedDriver Environmental permit, final and legally binding (Jan 2025) ● measured
↓ sign + · lag gate prerequisite
Responder Ren-Gas Tampere e-methane FID ○ projected
Mechanism An environmental permit is a legal prerequisite for construction of an industrial facility in Finland. Without it the project cannot proceed to FID.
Evidence, magnitude and the path to a testable form ▾
Ren-Gas Tampere FID requires committed offtake (resolved)
mechanism● ConjecturedDriver Gasum full-offtake agreement for the e-methane output ● measured
↓ sign + · lag gate prerequisite
Responder Ren-Gas Tampere e-methane FID ○ projected
Mechanism A committed buyer for the plant's output is required to underwrite project finance and reach FID. Gasum's full-offtake agreement covers the entire e-methane volume.
Evidence, magnitude and the path to a testable form ▾
Ren-Gas Tampere FID requires financing (resolved: EIB framework)
mechanism● ConjecturedDriver EIB EUR 230m approved framework loan ● measured
↓ sign + · lag gate prerequisite
Responder Ren-Gas Tampere e-methane FID ○ projected
Mechanism Project finance at a scale (~EUR 150-160m) that exceeds the sponsor's balance sheet. The EIB framework covers the debt; equity close and final disbursement conditions are the remaining steps.
Evidence, magnitude and the path to a testable form ▾
Ren-Gas Tampere FID requires feedstock electricity at viable cost (unresolved)
mechanism● ConjecturedDriver Electricity supply contract or PPA at a price that closes the e-methane business case ○ projected
↓ sign + · lag gate prerequisite
Responder Ren-Gas Tampere e-methane FID ○ projected
Mechanism Electrolysis at ~EUR 150-160m capex requires electricity below approximately 40-50 EUR/MWh to produce e-methane competitive with fossil gas. Finnish wholesale averages have been above this threshold since OL3 stabilised wholesale, but off-peak and PPA rates vary. No electricity supply contract has been publicly disclosed.
Evidence, magnitude and the path to a testable form ▾
The registered test, watched live: 2026 against the tripwire
● Under testThe decision rule was registered on 2026-07-07 and cannot be moved: it fires only if Terrafame's reported 2026 discharge lands outside 6,557–13,619 t (2025 was 10,088 t), and it then requires the measured flux to reach 28,359 t (upswing) or 18,906 t (downswing). The measured side below accumulates in public, from open SYKE data, before the operator's report publishes in spring 2027; that ordering is the independence of the test.
Sampling gate 3 of ≥5 sampled months so far; needs 2 more real sample months in the remaining 6(blue = SO4 sampled, grey = climatology fill; 2025 managed only 4)
No verdict is shown mid-year: a partial-year flux is not comparable to the annual thresholds, so until the year completes this panel reads accumulation and sampling adequacy only. Resolves by 2027-05-31, recorded publicly either way as ledger entry terrafame-lake-covariation-2026. Measured side refreshed monthly from SYKE VESLA and Hydrology (CC BY 4.0), snapshot 2026-07-10.
What this is, and is not
This is not a megatrend report and does not compete with one; orientation and accountability are different products. It is also not a claim to forecasting skill: with 18 open claims there is no sample to score, and there will be no accuracy percentages here until enough entries have resolved for the number to mean something. What the ledger demonstrates from day one is the mechanism: on this site, being wrong is possible, datable, and public. That property, not any single call, is the credibility argument, and it is the same property the rest of the site builds through per-claim citations and epistemic tiers (method).
Resolved entries stay on the page permanently. 5 resolved so far (all of them backtest worked examples, none counted).
Board sources and method (both boards)
- [1] US EIA, Short-Term Energy Outlook archive: 29 release vintages, 2018-2025 (public domain).
- [2] US EIA API v2: realized monthly WTI and Brent spot prices and US crude production (public domain).
- [3] Lopez, Pourjamal & Breyer (2025), 'An ex-post analysis of the IEA's World Energy Outlook', Renewable and Sustainable Energy Reviews 212:115371, CC BY 4.0; supplementary tables of WEO solar projections by vintage. Our parse reproduces the paper's own published 2020 error range (35.1-85.2% under, section 4.11) to the decimal when scored against its 741 GW denominator; the page's figures differ only by the OWID/IRENA realized series used here.
- [4] Our World in Data (IRENA-derived): world cumulative installed solar PV capacity, GW.
Every forecaster, including the baselines, uses only the history embedded in that STEO release file; STEO's own forecast is read from the same vintage. MAE per forecaster with n shown; skill is 1 - MAE/MAE(no-change); bias is the signed mean error. leaderboard_common restricts to rows where all core forecasters have values. Monte Carlo coverage is the share of actuals inside its 80% predictive band. Caveats: Prices (WTI, Brent) are effectively revision-immune; US crude production is scored against the current revised EIA series, not the vintage known at the time, so its errors mix forecast error with data revisions. The seasonal baseline only resolves for horizons up to 12 months and the rig model exists only for production; their own-sample MAE is not comparable to forecasters scored on all horizons. Use leaderboard_common or mae_by_horizon. Quarterly STEO vintages (plus 2025-01), a subsample of the monthly publication. Solar board: Each WEO vintage's CENTRAL scenario (New Policies for 2010-2018, STEPS from 2019) is scored on its world solar PV capacity projections at the milestone years that have since resolved. Signed percent error = (projected - realized) / realized; negative means the projection undershot reality. Only resolved pairs are scored; no forecaster-style bench applies here because the exhibit is bias, not relative accuracy. Caveats: IEA states its scenarios are conditional on policy, not forecasts. The central scenario is scored because it is the one read as the outlook by markets and governments; short-horizon (about five-year) errors are reported separately because policy in force at release largely determines a five-year build-out. Vintages before WEO 2010 are absent from the compilation (early editions reported solar inside 'other renewables'). Capacity definitions (DC vs AC nameplate) differ by a few percent between sources, immaterial next to the observed errors. WEO 2023-2025 are not in the compilation; extending the board to them needs manual extraction from the free WEO PDFs.