How Our Forecasts Work

1. What's observed vs forecast vs scenario

Every number on a Meridian page is one of four kinds:

• Observed — a measurement from an external agency at a known recent date. Population, sector employment, median income, named employers, AI exposure score, Consumer Stress, Aging Cliff. These are real measurements; Meridian sourced them.
• Forecast — a future-state estimate from our backtested projection model. The Conditions Profile dim values, the Projected Impact through 2031 panel. Always shown with an uncertainty band when our display contracts are followed.
• Scenario — a what-if number under a stated assumption set. The Scenario Explorer.
• Assumption — a model-internal value (a cascade weight, a ramp shape, an editorial impact estimate) that drives the forecasts but is itself not externally observed.

2. The forecast bands

We backtested our projection model from 2020 forward to 2026. For each of 9 geopolitical zones × 9 STEEPE dimensions × 3 target years (2022, 2024, 2026), we computed the projection from a 2020 baseline, the realized value, and the residual. That's 243 observation pairs.

Per-dimension Mean Absolute Error

6-year horizon. We report these honestly. Where you see a wide band on the Conditions Profile, that's the model telling you it's less confident there.

Forecast band math

For any forecast at year Y for zone Z and dimension D:

center         = projectSteeple(Z, Y)[D]
scaled_MAE     = MAE(D) × sqrt(yearsOut / 6)
probable_PI    = center ± 0.674 × scaled_MAE   (50% prediction interval)
plausible_PI   = center ± 1.60  × scaled_MAE   (89% prediction interval)
preferable_PI  = center ± 2.58  × scaled_MAE   (99% prediction interval)

Backtest interval coverage

• Outside-all-bands (realized fell outside the 99% PI): 1.2% (target 1%)
• Inside the preferable (99%) band: 98.8% (target 99%)

This supports the forecast-band display contract. It is not the same as a live public Brier accuracy claim; public Brier claims use the canonical Forecastable Model-only metric above and remain evidence-gated.

3. Cascade conventions (embedded priors)

When a scenario in our library encodes an effect, it triggers cascades to other dimensions. The conventions used by scenario authors are:

economic shock (−X) → social −0.6X, education −0.4X, political −0.3X, techBTC +0.5X
economic boom (+X)  → social +0.4X, education +0.3X
CBDC adoption (+X)  → techBTC +0.5X, political −0.4X
USD reserve loss    → techBTC +X, social(NA) −0.6X
Tech leap (+techAI) → education −0.5X, social variable

You should read these as the model's worldview, not as observed facts. Several encode debatable economic theories:

• economic shock → techBTC +0.5X. Assumes Bitcoin benefits from economic stress. There is some historical support (Cyprus 2013, Venezuela 2017-) and counter-evidence (March 2020 BTC sold off WITH equities; May 2022 stablecoin collapse).
• CBDC adoption → techBTC +0.5X. Assumes Bitcoin gains adoption when CBDCs are rolled out (as a privacy / sovereignty hedge). Plausible but not certain.
• USD reserve loss → techBTC +X. Assumes Bitcoin would be a primary beneficiary of USD reserve-status erosion. Gold and commodities and foreign currencies are at least as plausible.

We disclose these conventions explicitly. If you disagree with them, you should know they're in the model. A STEEPE v2 architecture (separate asset and policy dimensions) is in design where alternative paths exist.

4. Scenario library balance

~150 scenarios across 16 categories. Honest disclosure of representation by theme:

The library is heavily AI-focused, with substantial Bitcoin coverage, modest climate coverage, and minimal policy-response or energy-transition coverage. We are actively working on balance.

5. The three "cones" — now distinct entities

Historically Meridian called three different things "cone." We've cleaned this up. The three entities are:

• Forecast Bands. The 50% / 89% / 99% prediction intervals from the calibration above. The primary statistical-forecasting surface on the platform.
• Scenario Explorer. A what-if simulator. You select scenarios; the tool shows the resulting projected envelope. NOT a probabilistic forecast.
• Scenario Bucket. A categorical routing of scenarios into 4 narrative families (Most-likely, Aspirational, Plausible, Tail risk). Used to organize the library, not to generate forecasts.

6. Recommendations — what drives them

Meridian's recommendation engine reads county observed data (employment, demographics, income, education, infrastructure), AI exposure score, named employers, hand-authored LOCAL_ENRICHMENT per county, state incentive grounding, and CEDS records where available.

Of the recommendations on a county page, approximately 4% are driven by STEEPE scores (specifically, the ecological dimension when it crosses 65 triggers Climate Resilience). The remaining ~96% are driven by observed data.

7. What we know we don't yet do well

• County-level forecast calibration. We backtest at the continental scale. County pages inherit the bands.
• Joint-distribution modeling. Per-dim bands, not joint probabilities.
• Positive-vector outcomes. Our composite indices have been historically named with downside framing. We renamed them to neutral state names (Phase 1.1 of this sprint).
• Non-Gaussian shocks. Wars, pandemics, sudden regime changes. The calendar-anchored event overlay catches known past shocks; it cannot anticipate.
• Per-county scenario customization. The same scenario library applies to every county.
• Long-horizon dimensional independence. The radar visualization implies 9 independent dimensions; the underlying math produces only ~7 effective independent signals. See §11.
• Saturating ramp structure. Most cross-dim feedback factors use min(1, yearsOut/N) ramps. They produce calibrated 1-6 year forecasts but contribute zero motion past their saturation year. See §9.

9. Long-horizon saturation

Forecasts past 2030 plateau by design. Long-horizon values are asymptotic extrapolations, not signal-driven predictions.

Our STEEPE projection model uses bounded growth — headroom decay (100 − baseValue) / 100, ramp factors that saturate at 1.0 between 2029-2034, and a universal 0-100 clamp at every display layer. These mechanisms produce realistic short-horizon forecasts (calibrated to 1.2% outside-99%-PI over the 2020-2026 backtest, 243 observations) but bound long-horizon variance by construction.

In a 2026-2050 audit we generated trajectories for all 9 STEEPE dimensions × 10 geopolitical zones (90 series). 73% of those series plateau in some form by 2050; 31% plateau by 2030. Three dimensions plateau in every zone we tested: Social Trust, Quantum Readiness, and Education State.

What this means for you: when you scrub the timeline forward to 2040 or 2050, the values you see are not new forecasts. They are the asymptotes of the same mathematical curves you were already looking at at 2030. Treat them as directional ("this dim is still pressure-elevated at long horizon") rather than precise ("the model predicts 92 in 2050").

If a value sits at exactly 100 for many consecutive years (e.g. MENA Political State 2027-2050), the projection exceeded the 0-100 display range and was clamped. The clamped value is best read as "model registers maximum risk and cannot distinguish further" rather than as a precise forecast.

10. Input coverage horizon

Live inputs end around 2030. Beyond that, the model relies on bounded mathematical extrapolation.

The forecast engine can incorporate live signals — FRED unemployment + yield curve + repo data, ACLED conflict intensity, World Bank macro signals, prediction-market probabilities. These modulate per-dim drift rates and zone-adjustments.

Live signals are most informative at near-term horizons. They lose granularity progressively as the forecast moves past 2030 because the underlying data sources don't extend that far. By approximately 2035, live-input contribution drops to near-zero; the forecast becomes a pure function of the 2026 baseline + the per-dim mathematical structure.

What this means for you: 1-4 year decisions (PCAPP workforce transition, IRA infrastructure timing, sector-pivot programs) are where the forecast is most informative. 15+ year decisions (long-horizon liability matching, multi-decade capital planning) should treat the forecast as "the baseline if nothing else changes" — not as a probabilistic prediction.

11. Correlated dimensions

Education State and AI > AGI > ASI are highly correlated (r ≈ 0.88). Treat them as one indicator, not two.

The Education State function reads techAI directly: 55% of the destruction rate term is a function of the AI capability score. Zone-specific credential collapse, AGI-inflection, and ASI-extinction factors are added on top, but those are themselves correlated with year, so the resulting Education trajectory tracks techAI tightly.

Empirically, across all zones and all forecast years 2020-2050, the Pearson correlation between Education State and AI > AGI > ASI is 0.882. They are not independent forecasts.

Other moderately correlated pairs (r > 0.5):

• techAI ↔ techQuantum (r = 0.67)
• techAI ↔ techCBDC (r = 0.60)
• techAI ↔ techBTC (r = 0.59)
• Economic ↔ Political (r = 0.67) — this one is causally expected, not artifactual

What this means for you: when Education State and AI > AGI > ASI are both elevated on the radar, that is one signal corroborating itself, not two independent signals. The 9-dimension radar overstates the effective independent dimensionality of the forecast — the true effective count is closer to 6-7.

12. Score construction effects

STEEPE scores are 0-100 bounded by construction. Plateaus and ceilings can be score-construction artifacts, not signal.

Every STEEPE dimension is rendered on a 0-100 index. Projection functions can compute values above 100 internally; the display truncates them to 100. Values below 0 are clamped to 0.

This has three consequences:

• Plateaus at high values: when underlying drivers would produce 110, 120, or 150, the display sits at 100. The chairman sees a flat ceiling; the model "knew" the drivers were intensifying but cannot show it.
• Diminishing-returns visualization: many per-dim functions use headroom = (100 − baseValue) factors. These produce naturally asymptotic curves. A dim moving 87 → 92 over 14 years has slowed-but-not-stopped — but visually reads as "stopped."
• Cross-zone comparison can be misleading: a zone clamped at 100 looks similar to a zone genuinely converging to 100. The display does not distinguish.

What this means for you: a plateau in the display does NOT necessarily mean a plateau in real-world conditions. It can mean score-construction has hit its ceiling. Use clamped or plateaued readings as "this dimension is at maximum-pressure per the model" rather than "no further change expected." Where rigorous quantitative work requires unbounded readings, the underlying real-value inputs (industry shares, FRED macro signals, ACS demographics) are available via the OBSERVED chips on county pages.

13. Forecast tier definitions

Not every forecasting surface on Meridian carries the same epistemic weight. We classify each surface into one of four tiers and label it accordingly.

• ● Decision Grade. Live signal-driven, validated against ground truth where available, suitable for decisions within a stated horizon (typically ≤ 6 years). Examples: Economic state (FRED-wired), Action Items, Projected Impact through 2027, Industry Shares (ACS), Demographics.
• ◐ Directional Signal. Signal-informed and event-responsive. Useful as a compass to orient strategy. Not formally validated against ground truth. Treat as direction, not measurement. Examples: Political and Social state (≤ 2030 horizon), short-horizon Forecast Bands, STEEPE radar at near-horizon.
• ◇ Scenario Exploration. Generated from model-driven scenarios, not live signals. Useful for testing how the platform thinks about a possibility, not for committing to it. Examples: Scenario Explorer counterfactuals, PROJECTION_EVENTS overlay, near-horizon Ecological and Education trajectories.
• ◯ Research Surface. Research-grade output from an exploratory dimension or long-horizon projection. Subject to known structural limits (see §14). Shown for transparency in admin tooling, not as a recommendation. Examples: techQuantum, techCBDC, Anti-AI Resistance scenario, Wealth Inequality cone, long-horizon STEEPE radar (> 2030).

Each surface in the product wears its tier badge on the top-right of the component. Hover the badge for a one-sentence definition. The chairman walkthrough always identifies which tier is being discussed.

14. Research surface limitations

Research Surfaces (Tier ◯) carry specific, named structural limits. We disclose them rather than hide them.

In a 2026-06-14 elasticity audit (90 zone × dim trajectories, ±25 and ±50% baseline perturbation, 1,368 measurements), we found:

• Zero of 72 cross-dim pairs cross 5-point elasticity transmission. Cascade rules are documented in code (19+ effects in projectSteeple) but transmit too weakly to register as decision-relevant.
• Five of nine STEEPE dims converge to identical 2050 values across eight wildly-different scenarios. techCBDC=39 in 9/9 scenarios. techBTC=86, techQuantum=72, ecological=36, education=85 in 7/9 scenarios. These are predetermined, not signal-driven.
• 14 of 14 high-correlation pairs are synthetic aliases. Highest-r pair: techQuantum-education at r=+0.971 with zero cross-elasticity. The 9-dim radar overstates effective dim count; structurally there are ~4-5 independent dimensions, presented as 9.
• techQuantum is 100% formula-driven and shape-identical across all 10 tested zones. No live signal, no zone variation, no scenario sensitivity. It is a capability curve labeled as a dimension.
• Seven of nine STEEPE dims have zero live signal channels on the chairman path. Only Economic (full FRED wiring) and Social (partial UMCSENT wiring) flow through. Ecological has channels defined but unwired.

Implication: long-horizon forecasts (post-2030) and the tech-* sub-dims should be read as research output reflecting model structure, not as signal-driven predictions reflecting what the world is doing. The chairman view demotes these surfaces to opt-in or admin-only on this basis.

Source audits: docs/audits/forecast-reality/2026-06-14/ (Phases 1-10, including elasticity, counterfactual, plateau-stress, independence, narrative-dominance probes) and docs/audits/dimension-legitimacy/2026-06-14/ (10-phase ontology audit). Raw data + replay probes included.

15. How to verify our claims

For the rigorous reader, source paths:

• Forecast band math: src/lib/forecast-cones.ts + tests/fixtures/cone-band-mae.json
• Backtest script: scripts/cone-accuracy-backtest.ts
• Cascade conventions: src/lib/scenarios-library.ts lines 24-29
• Recommendation engine: src/lib/preferred-recommendations.ts
• Per-dim projection logic: src/lib/dim-projections.ts
• Scenario library: src/lib/scenarios/categories/*.ts
• Forecast Dynamics Investigation: docs/audits/forecast-dynamics/2026-06-14/
• Forecast Reality Audit (10 phases): docs/audits/forecast-reality/2026-06-14/
• Forecast Positioning (tier + visibility): docs/audits/forecast-positioning/2026-06-14/
• Dimension Legitimacy (ontology): docs/audits/dimension-legitimacy/2026-06-14/
• Trajectory probe (90 series × 31 yr): scripts/_probe-forecast-dynamics.mts