Beyond the Point Forecast

Methods white paper

Abstract

Passenger-vehicle registrations emerge from interacting state, OEM, fuel, infrastructure, policy and lifecycle dynamics. Autiqa's framework joins conventional market accounting with complexity-aware diagnostics while keeping causal claims proportional to evidence.

The method does not replace measurement with metaphor. It begins with reconciled registrations, makes system boundaries explicit, and adds tools only where they reveal concentration, dependency, adaptation or uncertainty that an aggregate growth rate conceals.

1. System boundaries and evidence tiers

The measured system is VAHAN passenger-vehicle registrations by state, OEM and fuel. Wholesale dispatch, bookings, model launches, policy and macroeconomic records are context sources, not substitutes for measured registrations.

• Measured: directly computed from canonical registrations.
• Corroborated: a measured pattern supported by an independent official or primary source.
• Hypothesis: a plausible mechanism with competing explanations still open.

The accounting identity comes first

Complexity tools are useful only after the market arithmetic closes. For each reporting period, the national passenger-vehicle total must reconcile to the sum of states and to independently grouped OEM and fuel cuts within declared tolerances. A contribution statement is always a change over a named comparison period, not a synonym for current volume.

This discipline prevents analytical novelty from hiding a basic inconsistency. It also makes revisions traceable: when a release changes, Autiqa can locate whether the movement came from a state, OEM, fuel, mapping rule or source correction rather than silently changing the story.

• Keep the release ID and cutoff with every derived dataset.
• Reconcile before ranking.
• Name both numerator and denominator in contribution claims.
• Preserve previous editions when the source revises.

Claims inherit the quality of their evidence

Measured claims are reproducible from registrations. Corroborated claims join a measured pattern to an independent primary source. Hypotheses propose a mechanism while keeping alternatives open. The tier belongs to the claim, not the paragraph: a report can contain measured volume, inferred structure and a hypothesis about cause on the same page.

A lifecycle event can explain timing only when its geography and category footprint plausibly match the measured change. A policy announcement can provide context without proving that registrations moved because of it. Confidence labels make that boundary visible to readers and reviewable by editors.

2. Concentration, diversity and specialisation

HHI captures whether volume is concentrated in a few categories. Shannon entropy is converted into an effective category count, answering how many equally sized categories would create the observed diversity. Location quotients compare a state's OEM or fuel share with the national mix, surfacing specialisation without declaring causation.

For shares pᵢ, HHI = Σpᵢ². Shannon entropy is H = −Σpᵢ ln pᵢ and the effective category count is exp(H). A state–category location quotient divides the category's state share by its national share. Each measure answers a structural question; none identifies a cause.

Decomposition locates the change

Aggregate growth is decomposed across state, OEM and fuel dimensions. Absolute change identifies which entities moved the national total; percentage change describes momentum relative to each entity's own base. Both are published because either can mislead alone.

The next step is intersection. A state–OEM or state–fuel edge shows where the increment occurred. Intersections are pruned by materiality so a small, spectacular percentage does not outrank a large, decision-relevant change. Small-base rules affect rankings, never visibility: excluded entities remain in appendices with a flag.

3. Dependency networks

State–OEM and state–fuel matrices are treated as weighted bipartite networks. Edge weights identify where aggregate growth depends on a narrow relationship. Network views are pruned to decision-relevant edges; dense hairballs are analysis artifacts, not reader interfaces.

A readable network is an editorial decision

The full bipartite matrix belongs in analysis, not automatically on the page. The reader view retains edges that are material by volume, contribution, specialisation or risk. It then orders them by the decision question: growth engines, concentrated dependencies or transition corridors.

Pruning changes presentation, not calculation. Totals are computed on the full matrix, while the figure states the threshold used for display. This prevents a visually dramatic hairball from implying that every connection is equally consequential.

• Calculate on the full graph.
• Display only decision-relevant edges.
• State the pruning threshold.
• Show concentration and fragility beside growth.

4. Dynamics, anomalies and persistence

Rank mobility describes competitive reordering. Seasonal rolling median/MAD flags robust deviations. CUSUM-type diagnostics identify sustained level shifts. No flag receives a causal label until persistence, revision history, correlated dimensions and competing explanations have been reviewed.

An anomaly is a case file, not a headline

The anomaly ledger records size, robust deviation, persistence, affected dimensions and data quality. A one-month spike can reflect demand, supply timing, fleet registration, reporting cadence or revision. The ledger keeps these alternatives attached to the signal.

Persistence changes the question. Repeated deviations in the same direction, accompanied by share or rank movement across related dimensions, justify structural review. Even then, the label is a shift flag rather than a causal verdict.

5. Diffusion without causal overreach

EV, CNG and hybrid adoption is traced across state clusters and time. Spatial correlation can indicate diffusion, shared infrastructure or common policy exposure; it cannot by itself identify which mechanism caused adoption.

Feedback loops describe hypotheses that can be tested

Vehicle transitions can reinforce themselves: installed infrastructure expands viable use cases; adoption improves infrastructure economics; product variety reduces perceived risk; higher confidence supports further adoption. A balancing loop can work in the opposite direction through congestion, charging queues, fuel availability or policy withdrawal.

The loop is not evidence by itself. Autiqa uses it to identify measurable implications—broader state participation, increasing fuel diversity, persistent specialisation or changing OEM participation—and then checks whether the data exhibit those implications.

• Name the reinforcing or balancing mechanism.
• Translate it into observable signals.
• Specify a competing loop or common cause.
• Withdraw the explanation when the expected signals fail.

6. Forecast ensembles and scenarios

Holt trend, linear trend and CAGR forecasts are evaluated through rolling historical cutoffs. Inverse out-of-sample RMSE weights the ensemble. Residual distributions provide horizon-specific 80% and 95% intervals. Only complete months train the models; partial months enter a separately labelled nowcast layer.

Backtests reproduce the information available at each historical cutoff. One-, two- and three-month errors are scored separately so a method that is useful next month does not automatically dominate the third month. WAPE is reported beside RMSE because percentage error becomes unstable for zero and near-zero state series.

• Base: weighted central path under recent structure.
• Upside: stronger breadth, transition participation and positive residual conditions.
• Downside: concentration failure, adverse revisions or persistent interval misses.
• Invalidation: definition changes or shocks outside the empirical residual regime.

Rolling-origin validation recreates past uncertainty

At each historical cutoff, the model sees only the months that would have been available then. It forecasts the next one, two and three months and stores the residuals by horizon. This avoids rewarding a model for information it could not have known.

Method weights are state- and horizon-specific. A seasonal method may dominate one state while a linear method performs better elsewhere. If a series is intermittent or the methods cannot produce a valid fit, a seasonal-median fallback is used and the state is marked Low confidence.

Intervals are empirical error translated into decision space

The 80% band describes the central range of rolling-origin residuals; the 95% band carries the outer range available from the backtest. Bands are constructed by horizon and remain ordered around the base path. They are not claims that uncertainty is normally distributed.

A wide interval is information about model and data limits. It can reflect volatility, a small or intermittent base, structural change or few comparable seasonal observations. The decision response is often optionality, not another decimal place.

• Publish base and both intervals together.
• Never hide an interval because it is inconveniently wide.
• Use WAPE to compare error with state scale.
• Treat repeated interval misses as a structural-review trigger.

7. Coherent state and national forecasts

Independent state forecasts rarely sum exactly to an independently estimated national total. Autiqa treats that disagreement as a hierarchy problem, not a rounding nuisance. The national ensemble remains the anchor; the gap is allocated across state paths in proportion to their backtest variance, allowing less certain series to absorb more of the reconciliation adjustment.

Base, 80% and 95% paths are reconciled separately. Deterministic integer allocation removes the final rounding difference. Assertions require every state interval to remain ordered and every monthly state sum to match its national counterpart exactly.

Scenarios must change observable watchlists

A scenario is not a percentage haircut around the base. It describes a coherent combination of participation, concentration, transition and residual conditions. Every scenario names the signals that would strengthen it, weaken it or make the underlying model irrelevant.

This creates accountability between editions. When the outlook changes, the revision ledger records whether new actuals, wider state participation, a structural flag or a source revision moved the forecast. The old path remains visible as a previous belief.

• Keep the base explicit.
• Tie upside and downside to measurable conditions.
• Publish invalidation criteria.
• Explain every material forecast revision.

8. Validation and limitations

Validation reconciles national, state, OEM and fuel totals; checks equal-elapsed fiscal comparisons; reports method errors and interval coverage; and audits every causal sentence. Forecast accuracy is assessed out of sample, not from in-sample fit.

The framework remains limited by registration definitions, revisions, missing model-level measurement, changing policy regimes and structural breaks not represented in history. Early-warning indicators are screening devices, not reliable clocks for transitions. A model can be internally coherent and still fail when the system boundary changes.

A report is a versioned evidence state

Every edition carries a source release, reporting cutoff, completeness status and publication version. Provisional reports are superseded, never silently overwritten. Living forecasts show the change from the previous edition and explain why the view moved.

The web reader and PDF consume the same manifest, figures and tables. This prevents a chart headline on the website from diverging from the downloadable document. Static PDFs remain addressable so a reader can reconstruct what was knowable at the time of a decision.

• One manifest for web and PDF.
• Stable slugs and versioned artifacts.
• Public access without entitlement checks.
• Release verification before publication.

9. Intellectual foundations

The framework is informed by W. Brian Arthur's account of complexity economics and increasing returns; the OECD's systemic-thinking approach to policy; and research on critical transitions. Early-warning indicators are used cautiously because empirical performance varies across systems and real-world forecastability is contested.

• W. Brian Arthur, Complexity and the Economy.
• OECD, Systemic Thinking for Policy Making.
• Scheffer et al., Early-warning signals for critical transitions.
• Empirical reassessments of early-warning signal reliability in complex systems.