Experimental forecast labPublic test · version 0.3
SLPA uses dated public data, known events and clear assumptions to build conditional planning ranges. These are not facts about the future, official forecasts or investment advice.
Five original ranges remain unchanged. Two new contracts extend the shadow ledger to merchandise trade and cruise arrivals. Shaded bands are planning ranges built from stated assumptions, not confidence intervals.
CSO's preliminary series reports 4.9% real growth in 2024. Official quarterly GDP currently stops at Q2 2025.
Two consecutive monthly stay-over readings outside the planning path
SLPA planning range. Lower edge: an energy-heavy year with weaker tourism. Upper edge: the IMF January 2026 forecast; the current IMF page shows 2.0%.
| Input | Value | Vintage | Status | Source |
|---|---|---|---|---|
| IMF current country projection | 2.0% | Accessed 9 August 2026 | External estimate | International Monetary Fund ↗ |
| IMF 2026 forecast | 2.3% | January 2026 | External estimate | International Monetary Fund ↗ |
| World Bank 2026 forecast | 1.9% | April 2026 | External estimate | World Bank ↗ |
| World Bank 2025 estimate | 1.3% | March 2026 | External estimate | World Bank ↗ |
The range is conditioned on the positive April signal and added winter airlift. It is not an official tourism forecast.
The EIA July outlook gives US$82 and the World Bank June outlook gives US$94. The vintage spread is retained rather than silently selecting one path.
No completion-rate assumption is made until public project execution data are available.
Use the official 2024 level as the last annual actual, place published forecasts beside it, and change the planning range only when tourism, quarterly output, energy or capital-execution signals materially move.
LimitDo not average conflicting 2025 estimates into a made-up actual.
Lag is measured from the end of the reference period to 9 August 2026. A recent publication can still describe an old period.
| Series | Reference period ends | Lag | State | Forecast effect |
|---|---|---|---|---|
| Annual GDP | 2024-12-31 | About 19 months | stale | Use 2024 as the last actual and show later figures as estimates or forecasts by vintage. |
| Quarterly GDP | 2025-06-30 | About 13 months | stale | Do not label a 2026 GDP estimate as an official nowcast. |
| Consumer Price Index | 2025-12-31 | About 7 months | aging | Use external energy and food drivers as scenarios, not as observed Saint Lucia inflation. |
| Labour Force Survey | 2026-03-31 | About 4 months | aging | Retain Q1 as a dated signal and do not extrapolate it as the current rate. |
| Stay-over arrivals | 2026-04-30 | About 3 months | aging | Use April as an early signal and wait for more months before moving the annual range. |
| Merchandise trade | 2025-12-31 | About 7 months | aging | Forecast annual magnitudes only and retain the published sign correction note. |
| Retail fuel-price cycle | Decision published 2026-08-03 | 6 days; prices effective through 23 August | current cycle | Show the effective-through date and never infer the next decision. |
| Reservoir and water-system operations | Dam elevation 2026-01-22; local service stress 2026-05 | About 6.5 months for elevation; about 3 months for the later local stress report | data gated | Monitor service stress. Abstain from a dam-level forecast. |
Six dated indicators show where the current evidence is leaning. A signal can trigger a review, but it cannot silently change a frozen forecast.
The early-year official signal supports the stay-over range, but four months cannot resolve the full-year path.
The updated external energy path keeps pressure on transport, electricity, food costs and subsidy choices. It is not a local price forecast.
The reading is official. A direction call is withheld while the 19th-to-21st-ICLS method transition and comparable 2026 quarters remain unresolved.
This is a budget plan, not an outturn. Revenue, capital delivery and support costs decide the realised path.
The gap narrowed from 2024, but higher energy and import-intensive investment could reverse part of that movement.
The basin outlook is not a Saint Lucia landfall forecast. CariCOF also keeps drought, humid heat and isolated flash flooding in the same planning window.
Revision rule: no indicator moves a range by itself. It must cross the published trigger, use a comparable definition and create a dated ledger entry.
Possibly, but only as soft data beside official indicators. The pilot will test observed language signals in shadow mode. It will not publish a national mood score.
No live index appears until source coverage, Kwéyòl performance, privacy safeguards and out-of-sample value are demonstrated.
Track reporting about prices, jobs, tourism, business activity and public services. Separate reporting tone from the event being reported.
Measure attention, expectations and uncertainty in legally accessible public discussion. Cap prolific sources and remove reposts.
Keep visitor-facing search and travel language separate from resident sources. Test whether it improves short-horizon arrival estimates.
Recent central-bank research finds that soft data can improve some short-horizon forecasts, but it can also create false alarms. The only useful question is whether a carefully governed signal adds real out-of-sample skill beyond official data.
These paths use RIPPLE-4 to follow effects and Sovereign Option Theory to set revision gates. They are not ranked by optimism.
Imported energy remains costly while government continues to smooth part of the retail shock.
Higher costs can weaken demand, tax receipts and the debt path
Positive early-2026 stay-over growth is followed by added British Airways winter capacity.
Local wages, procurement and taxes determine the national gain
The FY2026/27 Estimates plan a large programme while delivery capacity and financing remain constrained.
Operating costs, debt service and maintenance shape the future balance sheet
Rainfall can improve while catchment inflow, storage and distribution reliability remain weak.
Emergency works and lost activity create fiscal and economic pressure
Confirmed dates are kept separate from recurring windows and undated release watches. None tells us the result in advance.
The current official gasoline and diesel price is EC$17.25 per imperial gallon through 4 October; later decisions remain a publication watch.
Update the retail-price, subsidy and inflation mechanism. Do not infer the result before publication.
Regional credit, deposits and monetary conditions help test the domestic-demand and financing picture.
Admit only newly released series with a clear Saint Lucia reference period and vintage.
A new debt stock may narrow the published range, but only if its institutional boundary and GDP denominator are stated.
Reconcile the boundary before revising the debt card. A new number is not automatically a comparable number.
Canadian rates and the Canadian dollar can influence travel demand and spending from a key visitor market.
Treat the decision and exchange-rate response as source-market inputs, not as direct tourism outcomes.
Global food prices are an upstream input to Saint Lucia's imported food and household-cost path.
Use the index by category and with a lag. Do not treat it as observed local food inflation.
The EC dollar's fixed US-dollar rate transmits US monetary conditions into regional financing and exchange-rate competitiveness.
Track the rate decision and guidance as an external condition. Do not forecast a change in the EC dollar peg.
UK rates, growth and sterling can affect travel demand and the spending power of a major visitor market.
Use the decision as a UK-demand input, alongside bookings and airlift conversion.
A new regional release may change the output baseline and the denominator used in fiscal and debt ratios.
Freeze the release vintage, compare it with CSO definitions and revise only through a dated ledger entry.
The published schedule adds 51,000 seats across the 2026/27 winter season.
Treat the seats as capacity. Update arrivals only when bookings and observed passengers confirm conversion.
The fiscal and tourism risk calendar changes after the formal season, though hazards can occur outside it.
Retire the seasonal event flag, but retain actual recovery costs and water-system effects.
The outlook assumes severe Strait of Hormuz constraints through August and puts Brent near US$85 per barrel in Q3 and US$78 in Q4. This raises Saint Lucia's external energy-pressure screen.
Refresh the live imported-energy signal. Do not translate a global price path directly into a local pump, electricity or fiscal forecast.
NOAA assigns a 75% chance of a below-normal season and projects 7 to 13 named storms, 2 to 6 hurricanes and 0 to 2 major hurricanes. Most remaining activity is expected from August through October.
Use the outlook as basin-wide seasonal context, but do not convert it into a Saint Lucia landfall probability. One event can still dominate the year.
A single event can change output, tourism, water operations, public spending and debt at once.
Use actual watches, warnings and impact reports. Seasonal counts do not predict a Saint Lucia landfall.
The August 2026 adjustment is EC$0.274 per kWh and reflects fuel purchased in the prior month.
Resolve the next direction only after LUCELEC posts it. Do not infer the release day or map crude oil one for one into the surcharge.
Regional drought classification is a usable water-stress signal when reservoir operations are not published regularly.
Update the categorical water-pressure path, not a dam-level forecast.
The survey is intended to update poverty measures and CPI weights after data collection ended in April 2026.
Rebase the price and household-pressure modules only after CSO publishes results and methods.
New household and child indicators can change the social baseline used in policy scenarios.
Add the results as observed conditions. Do not infer completion from the July progress notice.
A boundary-aware reconstruction of the FY2026/27 budget plan, December 2025 public debt and selected separately audited public entities. It is not a consolidated sovereign balance sheet or an audit opinion.
Public debt stock. EC$5.132bn of central-government debt plus EC$308.3m of guaranteed debt. Guaranteed debt is already included, not added again.
Official Estimates envelope. The official financial plan includes fiscal-balance spending, principal repayment and refunds. It is not actual spending or the deficit.
Net revenue and grants. Positive revenue and grants after EC$10.014m in refunds. This is not cash already collected.
Planned overall deficit. Net revenue and grants minus fiscal-balance spending. Adding EC$140.0m of principal gives EC$352.4m in planned gross financing.
Government payables. The Economic and Social Review reports this stock outside public debt. The underlying creditor and ageing ledger is not published with the summary.
CIU cash and cash equivalents. CIU cash is a separate entity stock, not unrestricted Treasury cash and not a measure of total CIP wealth.
| Item | Value | Period | Status | Boundary note | Source |
|---|---|---|---|---|---|
| Nominal GDP | EC$7.049bn | 2024 | Official preliminary | Latest CSO annual current-price estimate, marked preliminary, and the scale denominator for historical ratios. | |
| Public debt stock | EC$5.440bn | 31 December 2025 | Official actual | EC$5.132bn of central-government debt plus EC$308.3m of guaranteed debt. Guaranteed debt is already included, not added again. | |
| Official Estimates envelope | EC$2.1888bn | FY2026/27 | Budget plan | The official financial plan includes fiscal-balance spending, principal repayment and refunds. It is not actual spending or the deficit. | |
| Net revenue and grants | EC$1.826bn | FY2026/27 plan | Budget plan | Positive revenue and grants after EC$10.014m in refunds. This is not cash already collected. | |
| Planned overall deficit | EC$212.4m | FY2026/27 plan | Budget plan | Net revenue and grants minus fiscal-balance spending. Adding EC$140.0m of principal gives EC$352.4m in planned gross financing. | |
| Foreign reserve cover | 3.6 months of imports | IMF 2026 projection | External estimate | An external-buffer estimate for the currency union context. It is not a central-government cash balance. | |
| Government payables | EC$20.4m | December 2025 | Official actual | The Economic and Social Review reports this stock outside public debt. The underlying creditor and ageing ledger is not published with the summary. | |
| CIU cash and cash equivalents | EC$259.9m | 31 March 2025 audited entity account | Official actual | CIU cash is a separate entity stock, not unrestricted Treasury cash and not a measure of total CIP wealth. |
Show revenue, primary spending, interest, capital execution, deficit, financing and payables together.
Keep airport, utility and other corporate debt and operating obligations visible outside the central-government box.
Separate programme revenue, programme cash, due-diligence fees and transfers to government.
Do not treat social-security assets as available central-government cash.
List guarantees, public-private partnerships and disaster facilities separately until called.
Do not subtract a few visible cash balances from comprehensive debt and call the result sovereign net worth.
Abstaining is a model output. Each blocked question has a specific evidence gate that can reopen it.
Table 2 in the official Q1 2026 report gives 12.1%, but Table 1 gives an incompatible 18.2% rate and different population totals. The annual 2025 web displays also conflict, and an official 20 April notice confirms a transition from 19th- to 21st-ICLS standards after a pilot.
Do not publish a point forecast for unemployment.
Restart rule: Begin with a directional scenario only after the historical and current series are methodologically comparable.
LUCELEC posts an August 2026 fuel cost adjustment of EC$0.274 per kWh, based on fuel purchased in the prior month. A complete machine-readable history was not located.
Do not publish a six-month point tariff.
Restart rule: Publish bill sensitivities first. Forecast a tariff only when the pass-through timing and full tariff stack can be reproduced.
The last direct John Compton Dam point found was about 332 feet on 22 January 2026. Later reporting describes severe local inflow stress, while CariCOF provides categorical drought evidence.
Do not forecast a reservoir elevation or national water-supply date.
Restart rule: Build a water-stress condition index before attempting an elevation forecast.
The tabled FY2024/25 report is described as EC$402.2m revenue, but applications, approvals, route mix and transfers can shift sharply.
Do not extend EC$402.2m as a trend line.
Restart rule: Forecast central-government transfers separately from programme revenue and cash.
Stay-over arrivals are available, while current visitor nights, spend, ownership and local procurement are incomplete.
Do not turn arrival counts into a dollar-value forecast.
Restart rule: Publish gross receipts, imported inputs and retained local value as separate lines.
Debt, budget flows, selected cash and some project liabilities are visible, but the asset and contingent-liability perimeter is incomplete.
Do not publish a sovereign net-worth number.
Restart rule: Expand the boundary in audited layers and publish reconciliation notes before calculating net worth.
No result counts unless the target, issued estimate, evidence cutoff and resolution rule were fixed before the outcome.
The first round was issued 9 August 2026. The trade and cruise expansion was issued 10 August 2026. Each used evidence available through its stated cutoff in Saint Lucia time. Later evidence may create a dated revision, but it cannot replace either original.
Issued 9 August 2026 · evidence cutoff 9 August 2026 · 5 contracts
| Forecast contract | Frozen planning band | Point estimate | Comparator | Status |
|---|---|---|---|---|
| SLPA-R1-2026-01-GDPReal GDP growth in 2026Calendar year 2026 | 0.5% to 2.3% | Withheld | IMF current country projection: 2.0% | open |
| SLPA-R1-2026-01-CPIAnnual-average inflation in 2026Calendar year 2026 | 1.5% to 3.5% | Withheld | World Bank projection: 2.1% | open |
| SLPA-R1-2026-01-STAYStay-over arrivals in 2026Calendar year 2026 | 430,900 to 452,300 | Withheld | No-change baseline: 426,676 arrivals | open |
| SLPA-R1-2026-01-FISCOverall fiscal deficit in 2026/27Fiscal year ending March 2027 | 2.5% to 3.5% of GDP | Withheld | IMF January 2026 projection excluding estimated disaster cost: 2.5% of GDP | open |
| SLPA-R1-2026-01-DEBTPublic debt at end-2026End of calendar year 2026 | 75% to 80% of GDP | Withheld | IMF January 2026 projection: 77.2% of GDP | open |
Issued 10 August 2026 · evidence cutoff 10 August 2026 · 2 contracts
| Forecast contract | Frozen planning band | Point estimate | Comparator | Status |
|---|---|---|---|---|
| SLPA-R1-2026-02-TRADEMerchandise trade gap in 2026Calendar year 2026 | EC$2.17bn to EC$2.48bn | Withheld | No-change component gap: EC$2.220bn | open |
| SLPA-R1-2026-02-CRUISECruise-passenger arrivals in 2026Calendar year 2026 | 668,100 to 768,300 | Withheld | No-change baseline: 668,086 arrivals | open |
Source grade A means a primary official record. Grade B means a multilateral estimate, research source or clearly labelled public-record reconstruction. Status and vintage still control how each figure may be used.
Official preliminary real-growth baseline. The table reports 4.9% growth in 2024 and marks the year preliminary.
Central Statistical Office of Saint Lucia ↗A · 2024 preliminaryOfficial preliminary annual baseline. The linked current-price table reports EC$7,049.1m nominal GDP in 2024.
Central Statistical Office of Saint Lucia ↗A · Through Q2 2025Latest public official quarterly output baseline located.
Central Statistical Office of Saint Lucia ↗A · Revised April 2026Official revised annual series reports 0.7% annual-average inflation in 2025.
Central Statistical Office of Saint Lucia ↗A · Data collection May 2025 to April 2026The survey is intended to update poverty measures and CPI weights. No results date is published.
Central Statistical Office of Saint Lucia ↗A · Q1 2026SLPA uses the 12.1% Table 2 rate. Table 1 reports an incompatible 18.2% rate and different population totals, and participation values vary within the report. CSO clarification is needed. The quarter is not interchangeable with the annual 2025 average.
Central Statistical Office of Saint Lucia ↗A · 2025 annualDetailed table reports 12.6% for 2025 while the CSO homepage displays a different headline value.
Central Statistical Office of Saint Lucia ↗A · 20 April 2026Confirms a planned transition from 19th- to 21st-ICLS standards after a pilot. It does not publish a bridge estimate or the first production quarter under the new method.
Government of Saint Lucia · Department of the Public Service ↗A · 2025 annualOfficial baseline for arrival episodes. The 2025 table reports 426,676 stay-over, 668,086 cruise and 41,349 yacht arrivals.
Central Statistical Office of Saint Lucia ↗A · April 2026Reports 40,752 April stay-over arrivals, up 8.5% year on year, 3.7% stay-over growth for January to April, 34% cruise growth and 44% yacht growth in April.
Saint Lucia Tourism Authority ↗A · Service begins 25 October 2026Adds 51,000 scheduled seats across winter 2026/27. Scheduled seats are not observed visitors.
Saint Lucia Tourism Authority via Government of Saint Lucia ↗A · 2025 annualSLPA derives the EC$2.220bn gap magnitude from components because some published balance signs are inconsistent.
Central Statistical Office of Saint Lucia ↗A · FY2026/27 budget estimatesPrimary budget book. It supports the detailed revenue, spending, deficit and financing reconstruction, but records plans rather than outturn.
Government of Saint Lucia, Department of Finance ↗A · FY2025/26 preliminary and year-end outlookLatest execution view located. Revenue is preliminary, spending includes a year-end outlook and payables are reported separately from debt.
Government of Saint Lucia, Department of Finance ↗A · Debt stock at 31 December 2025Reports EC$5.440bn in public debt, comprising EC$5.132bn central-government debt and EC$308.3m guaranteed debt, with instrument, maturity and refixing detail.
Government of Saint Lucia, Debt and Investment Unit ↗A · FY2026/27Lists a motion to adopt EC$2,188,765,900 in Estimates. The notice does not by itself establish that an Appropriation Act was enacted.
Parliament of Saint Lucia ↗A · FY2026/27Government summary reports payables of about EC$20m by December 2025.
Government of Saint Lucia ↗B · Published February 2026Reports a 74.5% public-debt ratio estimate for 2025.
Eastern Caribbean Central Bank ↗B · 14 January 2026Provides dated macroeconomic and fiscal estimates and forecasts. Keep its vintage visible.
International Monetary Fund ↗B · Accessed 9 August 2026The current page displayed 2.0% real GDP growth and 2.5% consumer-price growth for 2026. The access date is frozen because the page can update.
International Monetary Fund ↗B · April 2026 macro outlookEstimates 2025 growth at 1.3%, 2026 growth at 1.9%, FY2025/26 debt at 77.9% of GDP and 2026 inflation at 2.1%.
World Bank ↗B · 11 June 2026External scenario assumes Brent averages US$94 in 2026. It is not a Saint Lucia price forecast.
World Bank ↗A · 7 July 2026 outlookForecasts Brent at US$82 per barrel on average in 2026. This is an external energy path, not a Saint Lucia fuel-price forecast.
US Energy Information Administration ↗A · Released 7 August 2026The index was 131.1, up 0.6% month to month and 1.0% year to year, with mixed category movements. It is not observed Saint Lucia CPI.
Food and Agriculture Organization of the United Nations ↗B · April 2025Documents collection and compilation gaps affecting CIP, goods, travel and direct-investment statistics.
International Monetary Fund ↗A · 3 to 23 August 2026Gasoline and diesel were EC$16.75 per imperial gallon for 3 to 23 August. A successor notice following the scheduled 24 August review had not been located at the 28 August check.
Government of Saint Lucia ↗A · 22 January 2026Reports John Compton Dam near 332 feet and below spillway at that date.
Water and Sewerage Company Inc. ↗A · May 2026Reports local demand of 1.2m gallons per day and inflow at 20% of capacity. This is not a national reservoir reading.
Water and Sewerage Company Inc. ↗B · 2026 seasonal outlookRegional outlook indicates long-term drought could persist in Saint Lucia while flood risk also remains possible.
Caribbean Institute for Meteorology and Hydrology ↗A · 6 August 2026Assigns 75% probability to a below-normal basin season, 20% to near-normal and 5% to above-normal. The seasonal basin outlook does not estimate a Saint Lucia landfall probability.
NOAA Climate Prediction Center ↗A · Checked 12 August 2026Historical archive link. NOAA's live outlook page carries the 6 August 2026 update used in the current signal room.
NOAA Climate Prediction Center ↗A · Recurring Atlantic seasonUse for the seasonal calendar only. It does not predict a Saint Lucia landfall.
NOAA National Hurricane Center ↗A · Year ended 31 March 2025Audited entity account. CIU cash, liabilities and flows remain separate from a consolidated central-government balance sheet.
Citizenship by Investment Unit, Saint Lucia ↗B · FY2024/25Use as reported public-record detail until the primary annual-report PDF is directly inspectable.
Public reporting based on the FY2024/25 tabled report ↗A · 2026 scheduleOfficial dates for the global energy outlook used as an external input, including 11 August 2026.
US Energy Information Administration ↗A · Released 11 August 2026 · forecast completed 6 AugustForecasts Brent near US$85 per barrel in Q3 2026, US$78 in Q4 and US$69 on average in 2027. It is an external scenario, not a Saint Lucia retail-price forecast.
US Energy Information Administration ↗A · Accessed 9 August 2026Official regional release dates. Product dates do not guarantee a same-day CSO webpage update.
Eastern Caribbean Central Bank ↗A · 2026 schedulePrimary calendar for a major Saint Lucia visitor-source market.
Bank of Canada ↗A · 2026 scheduleGlobal upstream food-price signal. It is not a local CPI observation.
Food and Agriculture Organization of the United Nations ↗A · 2026 schedulePrimary schedule for US policy decisions relevant to the EC dollar's fixed US-dollar rate.
US Federal Reserve ↗A · 2026 schedulePrimary schedule for a major Saint Lucia visitor-source market.
Bank of England ↗A · August 2026Official page reports an August 2026 fuel cost adjustment of EC$0.274 per kWh.
Saint Lucia Electricity Services Limited ↗A · Current method noteExplains that the monthly adjustment reflects fuel purchased in the prior month.
Saint Lucia Electricity Services Limited ↗A · 30 July 2026Regional guidance places Saint Lucia’s long-term drought at an imminent end-November horizon while retaining humid-heat and isolated flood risk. It is not a national alert or reservoir-level forecast.
Caribbean Institute for Meteorology and Hydrology ↗A · July 2026 progress noticeA progress notice, not a results release or a dated publication promise.
Central Statistical Office of Saint Lucia ↗Public method
The lab estimates possible paths. It does not report facts about the future. Every forecast carries an evidence cutoff, a horizon, a resolution rule and a score after the outcome is known.
A confirmed date, rule or scheduled release. No probability is added.
An estimate of the present or recent past before an official result is released.
A dated hard or soft indicator that may trigger review. It does not change a forecast by itself.
A dated probability or numerical range for a defined future outcome.
An estimate that applies only if the named assumptions hold.
A coherent path used to test policy. It is not automatically the most likely path.
Fix the unit, geography, horizon, evidence cutoff and timezone, authoritative resolver, first-release or revised outcome, fallback and void rule before estimating anything.
Start with a reference-class base rate, last observation, seasonal naive path, recent trend and relevant official consensus. Compare the same target, horizon and issue date.
A human estimate is recorded before the model aggregate is shown. LLM runs are treated as one related model family, not a crowd of independent experts.
Begin with equal or median combinations. Weights can change only after rolling tests. Complements, exhaustive outcomes, nested events and numerical quantiles must remain coherent.
The original stays visible. Each update records the new evidence and numerical reason. Initial, fixed lead-time and latest eligible estimates are scored separately.
Mean squared distance between the stated probability and the outcome. Lower is better. A constant 50 percent forecast scores 0.25.
mean (probability − outcome)²Test-set mean absolute error divided by the mean absolute in-sample naive or seasonal-naive change. Below 1 means error is smaller than that scaling error. Signed mean error shows high or low bias.
test MAE ÷ in-sample naive scaleDeclared numerical distributions use weighted interval score. Nominal coverage and width are reported beside it. Current R1 bands have no nominal coverage, so inclusion is descriptive only.
proper interval loss + diagnosticsPerformance is compared with matched baselines. The formula below applies only to paired, positive, lower-is-better losses such as Brier score, not to bias or coverage.
1 − model score ÷ baseline scoreDefinitions, methods and sample cards.
Dated forecasts accumulate without shaping editorial claims.
Results and failures become public. No skill claim is made.
Sustained out-of-sample results beat declared baselines with useful coverage.
R1 deliberately withholds point estimates and does not assign probability coverage to its planning bands. The lab will not claim skill from band inclusion or a few successes. Later performance remains separated by target and horizon, with uncertainty around every score.
Prospective unresolved questions limit answer leakage. Expert forecasters outperformed the leading LLM in the initial study.
Open source 02 · Peer-reviewed · NeurIPS 2024Dated retrieval and ensembling improved machine forecasts. A validation-chosen human and machine aggregate performed best on this binary-event test.
Open source 03 · Peer-reviewed · ICLR 2026Revised webpages, unreliable date filters, question selection and leaked market information can make retrospective AI forecasts look better than they were.
Open source 04 · Peer-reviewed · ACM TiiS 2025A preregistered study found useful human-assistant gains on six numerical questions. The narrow task set and outlier sensitivity limit generalisation.
Open source 05 · Emerging research · 2026 preprintTested models produced intervals that were too narrow. None reached the requested 90 percent coverage rate.
Open source 06 · Emerging research · 2026 preprintEvolveCast finds model updates can be inconsistent or too conservative. Retrieval alone is not a validated belief-updating rule.
Open source 07 · Peer-reviewed · NeurIPS 2024Across the tested methods and datasets, removing or replacing the pretrained language model often improved results. Numerical engines must earn their place in backtests.
Open source 08 · Peer-reviewed · International Journal of ForecastingAcross 100,000 series, combinations and hybrid methods dominated the strongest entries. Simple baselines remained essential.
Open source 09 · Peer-reviewed · Psychological ScienceProbability training, reference classes, team collaboration and selection of previously strong forecasters improved judgmental performance.
Open source 10 · Peer-reviewed · Journal of the American Statistical AssociationProper scores reward honest probabilities and intervals that balance coverage with useful precision.
Open source 11 · Peer-reviewed · PLOS Computational BiologyWeighted interval score combines calibration and sharpness for numerical distributions. Coverage and width remain useful separate diagnostics.
Open source 12 · Peer-reviewed · International Journal of ForecastingMASE scales test error by the mean in-sample naive or seasonal-naive change, enabling comparison across numerical series.
Open source 13 · Peer-reviewed · International Journal of ForecastingRolling forecast origins and repeated test periods provide a stronger test than one convenient historical holdout.
Open source 14 · Peer-reviewed · Journal of EconometricsHistorical tests must use the information available at the time. Later revisions can otherwise leak into apparent performance.
Open source 15 · Peer-reviewed · ICLR 2025Forecasts should pass coherence checks: complements and exhaustive outcomes must sum correctly, nested events must be monotone and quantiles must not cross.
Open source 16 · Official policy guidance · 2025Scenarios explore plausible paths, signposts and policy options. They are not automatically forecasts or mutually exclusive outcomes.
Open source 17 · Central-bank research · 2026Soft indicators improved some short-horizon models but also produced false alarms. They should be tested beside hard data, not used alone.
Open source 18 · Central-bank research · BIS IFCNews tone helped selected short-horizon targets but did not improve most broader output components. Usefulness must be target-specific.
Open source 19 · Peer-reviewed · NAACL 2024Creole-language performance cannot be assumed from English models. Local annotation and language-specific evaluation are required.
Open source 20 · Official analytics guidance · United NationsDigital-data projects should apply privacy-by-design and assess representativeness, manipulation risk and potential harm across the full data lifecycle.
Open sourceThis public test remains experimental. It will not guide wider SLPA analysis until it has a credible record on forecasts made before the results were known.
Return to the official Data Observatory