Anchored Horizon — Capabilities

Forecasts finance and producers
can actually plan against.

Anchored Horizon turns comparable-game evidence into distribution-first commercial belief—honest uncertainty, time-scoped horizons, scenarios you can challenge, and live updates as signals arrive—through the BeliefSession, API, MCP, and A2A.

  1. 01

    Plan against a distribution

    Replace comps decks and single-number forecasts with posterior predictive belief—bands, targets, hit chance, and evidence strength you can defend.

    Predictive belief

    • Posterior predictive draws (not a point estimate)
    • Planning quantiles (p10–p90 and custom bands)
    • Hit probability and P(revenue ≥ target)
    • Log1p modeling with proper back-transform to currency
    • FX normalization and deflation to a planning currency
    • Named forecast profiles (12-month, 24-month, as-of, mixture tiers)
    • Batch multi-title forecast runs
    • BeliefSession question scope: outcome, horizon, and anchor
    • Density and CDF views for distribution inspection
    • Pins, planning bands, and scenario forks in-session

    Evidence strength

    • ESS (Kish) from composite comp weights
    • ESS-scaled prior width and power-prior discounting
    • Explicit “unreliable when” gates (low ESS, sparse comps, missing embeddings)
    • Prior vs features vs groups vs tail explanation
    • Comp influence ranking on every forecast
    • Lineage artifacts for audit and reproducibility
  2. 02

    Anchor in comparable games

    Priors come from titles that are similar in game space—weighted, overridable, and explainable—not from a hand-picked deck.

    Comp retrieval & weighting

    • Latent Spark similarity retrieval (asymmetric KL over Gaussian embeddings)
    • Top-k comps with query fuzziness and diversity quotas
    • Composite weights: similarity × recency × quality
    • Regime-aware recency half-life for aging comps
    • Multi-signal quality scores (reviews, CCU, rank, profile signals)
    • Standalone modes when Latent Spark is not connected
    • Comp-set stability index for analyst review

    Analyst control

    • Include, exclude, and reweight comps before prior construction
    • Explainable prior construction (comps inform priors only—no double-count)
    • Feature-conditioned likelihood for the subject title
    • Hierarchical hit-mixture model (base vs hit regimes)
    • Constrained mixture parameterization to prevent label switching
    • Forecast explain: what drove the outlook and how much
  3. 03

    Horizons, trajectories & time

    Every forecast is a time-scoped belief—windowed horizons, live signal trajectories, and multi-horizon paths for planning—not a timeless spreadsheet cell.

    Horizon-native forecasting

    • Horizon scopes: 12-month, 24-month, and custom month windows
    • Anchors: launch, as-of date, and fiscal year end
    • Scoped outcome keys per entity, horizon, and anchor
    • Multi-horizon path forecasts for strategic pathway planning
    • Path-horizon bundles (e.g. 3/6/12/24 or 6/12/24/36 months)
    • Temporal training windows and rolling-origin backtests
    • Regime-aware aging of comparable-game evidence

    Live trajectories

    • Time-stamped signal ingest (wishlist, engagement, revenue)
    • Trajectory features: slope, momentum, volatility, wishlist velocity
    • Pre-launch and post-launch trajectory modes
    • Signal panels that keep features current as the market moves
    • Observation scoring against frozen forecast runs over time
    • Cohort calibration drift checks across time windows
  4. 04

    Scenarios & inverse planning

    Ask “what if,” solve for the input that hits a plan target, and keep causal honesty when analysts move the knobs.

    Counterfactuals & compare

    • Feature-override scenarios that re-forecast under new assumptions
    • Ghost forks and side-by-side scenario compare
    • Net-revenue waterfalls (platform, publisher, deductions) on draws
    • Causal role labeling (intervention, mediator, conditioning, collider)
    • Non-blocking causal validation on scenario draws

    Solve for the decision

    • Inverse constraint solve (exceedance, median, or quantile targets)
    • Wishlist- and pin-style solves (“what must be true for 75% chance of ≥ $X”)
    • Scenario inverse APIs for agents and UI
    • Planning pins that stay attached as belief updates
    • Exportable scenario packs for finance and leadership reviews
  5. 05

    Calibration, surprise & lifecycle

    Keep forecasts decision-grade after they leave the lab—promotion gates, surprise scoring, and governed refit.

    Trust & calibration

    • Rolling-origin CRPS, log score, coverage, and PIT diagnostics
    • Per-horizon calibration segments
    • Candidate → active model promotion with statistical gates
    • R-hat / ESS / coverage thresholds before release
    • Rollback to prior active models
    • Diagnostickit promotion gates for safe operation

    When reality lands

    • Score realized outcomes against a frozen forecast_run_id
    • Calibration vs structural surprise classification
    • No silent prior override when observations arrive
    • Post-forecast drift detection
    • Auto-refit policy with cooldown and caps
    • Sensitivity and upside/downside partition explainers
  6. 06

    Portfolio, OIE & platform

    Title-level belief joins portfolio risk and the Iridae loop—or runs stand-alone when peers are not connected.

    Portfolio & peers

    • Joint portfolio draws across titles
    • Independent or Gaussian-copula correlation assumptions
    • VaR-style portfolio risk summaries
    • Subtle Beacon experiment-posterior ingest (promotion-gated)
    • Patient Cartographer tactic distributions and path-horizon payloads
    • Latent Spark comp retrieval and structural-surprise notify
    • Fail-soft mesh when peers are unavailable (wider, weaker priors)

    Surfaces & governance

    • Headless capability API (AgenticKit-only public surface)
    • OpenAPI, MCP, and A2A projections
    • BeliefSession for analysts; agents under the same contracts
    • Tenant isolation and role-based access (analyst / operator / admin)
    • Lineage chain for forecast runs and promotions
    • Optional commerce metering and customer-success hooks
    • Hosted multi-tenant SaaS

Predictive belief

  • Posterior predictive draws (not a point estimate)
  • Planning quantiles (p10–p90 and custom bands)
  • Hit probability and P(revenue ≥ target)
  • Log1p modeling with proper back-transform to currency
  • FX normalization and deflation to a planning currency
  • Named forecast profiles (12-month, 24-month, as-of, mixture tiers)
  • Batch multi-title forecast runs
  • BeliefSession question scope: outcome, horizon, and anchor
  • Density and CDF views for distribution inspection
  • Pins, planning bands, and scenario forks in-session

Evidence strength

  • ESS (Kish) from composite comp weights
  • ESS-scaled prior width and power-prior discounting
  • Explicit “unreliable when” gates (low ESS, sparse comps, missing embeddings)
  • Prior vs features vs groups vs tail explanation
  • Comp influence ranking on every forecast
  • Lineage artifacts for audit and reproducibility

Comp retrieval & weighting

  • Latent Spark similarity retrieval (asymmetric KL over Gaussian embeddings)
  • Top-k comps with query fuzziness and diversity quotas
  • Composite weights: similarity × recency × quality
  • Regime-aware recency half-life for aging comps
  • Multi-signal quality scores (reviews, CCU, rank, profile signals)
  • Standalone modes when Latent Spark is not connected
  • Comp-set stability index for analyst review

Analyst control

  • Include, exclude, and reweight comps before prior construction
  • Explainable prior construction (comps inform priors only—no double-count)
  • Feature-conditioned likelihood for the subject title
  • Hierarchical hit-mixture model (base vs hit regimes)
  • Constrained mixture parameterization to prevent label switching
  • Forecast explain: what drove the outlook and how much

Horizon-native forecasting

  • Horizon scopes: 12-month, 24-month, and custom month windows
  • Anchors: launch, as-of date, and fiscal year end
  • Scoped outcome keys per entity, horizon, and anchor
  • Multi-horizon path forecasts for strategic pathway planning
  • Path-horizon bundles (e.g. 3/6/12/24 or 6/12/24/36 months)
  • Temporal training windows and rolling-origin backtests
  • Regime-aware aging of comparable-game evidence

Live trajectories

  • Time-stamped signal ingest (wishlist, engagement, revenue)
  • Trajectory features: slope, momentum, volatility, wishlist velocity
  • Pre-launch and post-launch trajectory modes
  • Signal panels that keep features current as the market moves
  • Observation scoring against frozen forecast runs over time
  • Cohort calibration drift checks across time windows

Counterfactuals & compare

  • Feature-override scenarios that re-forecast under new assumptions
  • Ghost forks and side-by-side scenario compare
  • Net-revenue waterfalls (platform, publisher, deductions) on draws
  • Causal role labeling (intervention, mediator, conditioning, collider)
  • Non-blocking causal validation on scenario draws

Solve for the decision

  • Inverse constraint solve (exceedance, median, or quantile targets)
  • Wishlist- and pin-style solves (“what must be true for 75% chance of ≥ $X”)
  • Scenario inverse APIs for agents and UI
  • Planning pins that stay attached as belief updates
  • Exportable scenario packs for finance and leadership reviews

Trust & calibration

  • Rolling-origin CRPS, log score, coverage, and PIT diagnostics
  • Per-horizon calibration segments
  • Candidate → active model promotion with statistical gates
  • R-hat / ESS / coverage thresholds before release
  • Rollback to prior active models
  • Diagnostickit promotion gates for safe operation

When reality lands

  • Score realized outcomes against a frozen forecast_run_id
  • Calibration vs structural surprise classification
  • No silent prior override when observations arrive
  • Post-forecast drift detection
  • Auto-refit policy with cooldown and caps
  • Sensitivity and upside/downside partition explainers

Portfolio & peers

  • Joint portfolio draws across titles
  • Independent or Gaussian-copula correlation assumptions
  • VaR-style portfolio risk summaries
  • Subtle Beacon experiment-posterior ingest (promotion-gated)
  • Patient Cartographer tactic distributions and path-horizon payloads
  • Latent Spark comp retrieval and structural-surprise notify
  • Fail-soft mesh when peers are unavailable (wider, weaker priors)

Surfaces & governance

  • Headless capability API (AgenticKit-only public surface)
  • OpenAPI, MCP, and A2A projections
  • BeliefSession for analysts; agents under the same contracts
  • Tenant isolation and role-based access (analyst / operator / admin)
  • Lineage chain for forecast runs and promotions
  • Optional commerce metering and customer-success hooks
  • Hosted multi-tenant SaaS

See Anchored Horizon in action.

Explore how your studio can ground forecasts in comps, model the full range of outcomes, and keep commercial expectations current.

Book a demo