A Forecast Is a Field of Possibilities, Not a Number

Essay

A point hides the decision landscape

A single forecast compresses alternative futures into one coordinate. That can help budgeting, but it can also erase the range of conditions under which a decision succeeds or fails. The better object is a field: a base path, credible alternatives, uncertainty bands and signals that move probability between them.

A forecast becomes useful when it changes a decision. The question is not whether the base number will be exactly right; it is whether a plan survives plausible error, which signals justify moving resources, and which conditions make the original model irrelevant.

A point estimate invites a point response: order this many vehicles, reserve this much capacity, expect this share. Real decisions are rarely so smooth. They have thresholds, fixed costs, lead times and asymmetric consequences. Missing demand by ten percent may create a manageable stock imbalance; building irreversible capacity for an upside case that never arrives can impair years of returns.

Intervals translate model error into decision space. The central range describes outcomes that routine planning should survive. The wider range supports resilience questions: liquidity, supply flexibility and downside protection. Neither band removes uncertainty. They make its scale visible and prevent a decimal-heavy base case from monopolising attention.

Scenarios add structure to the range. An upside path should describe conditions such as broader state participation, improving transition diversity and positive residual momentum. A downside path should name concentration failures, adverse revisions or persistent misses. Percentages without mechanisms are sensitivity tests, not scenarios.

The field also needs an edge. Every model has conditions under which it ceases to be the right representation: a definition change, major policy shock, supply interruption or residual outside historical experience. Naming invalidation criteria is more useful than stretching an old interval around a new system.

The forecast date is as important as the value. A path issued before a launch, policy change or source revision belongs to a different evidence state from one issued afterward. Stable URLs and versioned artifacts allow decisions to be evaluated against what was knowable, not against a forecast silently refreshed with later information.

Markets remember how they arrived

Vehicle markets are path-dependent. A charging network changes which products are viable; product availability changes adoption; adoption changes infrastructure economics. OEM share today affects dealer reach and service confidence tomorrow. These feedbacks mean the same shock can produce different outcomes depending on the system's prior state.

Path dependence means that history influences the menu of future options. A state with established CNG infrastructure can absorb new compatible models faster than an otherwise similar state. An OEM with a dense service network can convert a launch into registrations more quickly. These advantages are produced by accumulated relationships, not only current-month demand.

Positive feedback can widen initial differences. More adoption improves utilisation, which supports infrastructure, which reduces perceived risk and attracts more adoption. Balancing feedback can interrupt the loop through congestion, constrained supply, declining incentives or poor service experience. A forecast that extrapolates one recent slope without representing these mechanisms can appear accurate until the regime changes.

This does not mean every market story needs a complex simulation. It means model selection should respect the behaviour visible in the data. Seasonal methods, linear trends and growth-rate models encode different assumptions. Rolling-origin testing reveals which assumptions have survived comparable historical cutoffs for each state and horizon.

The practical consequence is humility about transfer. A method that performs well nationally may be weak in an intermittent territory. A model that wins at one month may drift at three. Forecast governance should preserve those differences instead of forcing every geography into one elegant equation.

State interactions can also transmit shocks. A supply constraint at one OEM can redirect demand toward competitors unevenly across dealer networks; an infrastructure bottleneck can slow one fuel pathway while leaving the national total intact. Dependency maps help identify where a local miss could become a broader forecast error.

Forecast exhibit

Three horizons, one coherent geography

Autiqa forecasts each state across the next three months, evaluates each method at the same one-, two- and three-month horizons, and reconciles those state paths to an independently modelled national total. Without reconciliation, a national chart and its state appendix could disagree while both appeared precise.

The current national three-month path totals 12,88,640 registrations. The state appendix allocates that total exactly, while state-specific WAPE determines how confidently each local path should be used.

Coherence is an accounting requirement, not a cosmetic adjustment. Independent state models answer local histories; the national ensemble answers the aggregate series. Their sums will differ because estimation error is unavoidable. Reconciliation allocates that gap according to uncertainty and then applies deterministic rounding so every published integer closes exactly.

The interval paths are reconciled separately. This matters because simply scaling a base forecast can invert or compress uncertainty for small states. The published lower, base and upper paths must retain their ordering at every horizon while their state sums match the corresponding national path.

Ensembles discipline model confidence

Holt, linear and CAGR models fail differently. Rolling-origin backtests reveal those failure patterns. Weighting them by historical error is not a claim that the ensemble knows the future; it is a way to avoid letting one convenient story dominate without evidence.

Model diversity is useful only when performance is measured out of sample. At each historical cutoff, the method sees no future observations. Errors are stored separately for one, two and three months because useful short-horizon behaviour may not persist. Weights are then derived from those errors, invalid fits are removed and the remainder is renormalised.

The ensemble is not automatically superior in every case. Sparse states may offer too little stable history for trend or seasonal estimates. A disclosed seasonal-median fallback can be more honest than a fragile fitted curve. The state remains in the publication, receives Low confidence and carries a wider decision margin.

WAPE and RMSE answer complementary questions. RMSE preserves the scale of registration misses and penalises larger errors. WAPE relates absolute error to observed volume without the instability that MAPE creates near zero. Neither score is a guarantee; together they reveal whether a visually plausible line has earned operational trust.

Coverage is also scored. If an 80 percent interval contains far fewer than roughly eight in ten backtest outcomes, the band is overconfident. If it contains nearly everything, it may be too broad to guide action. Forecast quality includes the calibration of uncertainty, not only closeness of the central line.

Weighting should remain horizon-specific. The model that best anticipates next month may converge too quickly or drift too far by the third. Publishing one blended weight for all horizons would conceal that behaviour. Separate scores make the forecast field more coherent even when they reduce the simplicity of the method description.

• Use only complete months for training.
• Publish method-level error and interval coverage.
• Widen uncertainty with horizon.
• Track revisions as changed beliefs.

Counterargument

Scenarios can become a way to avoid choosing

A field of possibilities is not permission to publish every imaginable outcome. Too many scenarios dilute accountability. Useful scenarios are few, mutually intelligible and connected to observable signals.

The base path should remain explicit. Upside and downside should describe coherent conditions, not arbitrary percentages. An invalidation condition must say when the model is no longer the right frame. Uncertainty clarifies a decision only when it has boundaries.

Scenario proliferation often begins with good intentions. Teams add one path for every stakeholder concern until no outcome can falsify the report. The result is comprehensive but inert. A useful field remains deliberately small: a base case, bounded upside and downside, and a clear condition that invalidates the frame.

There is also a danger of outsourcing judgement to intervals. A model can quantify historical residuals but cannot know the strategic cost of a miss. Leaders still have to decide which errors are tolerable, which commitments are reversible and which signals deserve action. Forecasting organises that judgement; it does not replace it.

Accountability comes from revision history. When the base path changes, the report should attribute the movement to new actuals, model reweighting, a source revision or a scenario assumption. Quietly replacing last month's line destroys the learning value of the system and makes every forecast look better in retrospect.

Choice is therefore preserved by boundaries. Each scenario has a mechanism, observable watchlist and action implication. If two paths lead to the same decision, they do not need separate names. If a decision changes only at a threshold, that threshold should be visible on the page.

Some decisions still require a single planning number. The base path serves that role, provided its provenance and uncertainty travel with it. The argument is not to ban point estimates; it is to prevent the point from erasing the conditions, error history and alternatives that determine how firmly it should guide action.

Turn the interval into a decision rule

A robust plan maps actions to the forecast field. Capacity committed inside the 80% range can be treated differently from irreversible investment that only works near the upside edge. A state with a wide interval may deserve optionality; a state with a narrow but declining path may demand a strategic response rather than more forecasting.

Inventory decisions can be staged against horizon and confidence. Near-term base demand may justify normal allocation, while the upper 80 percent path informs flexible replenishment. The 95 percent range is better suited to resilience planning than to sales targets. Using one number for all three purposes confuses expected demand with risk capacity.

Geographic uncertainty should affect commitment design. A large state with a narrow interval may support fixed allocation. A fast-growing state with a wide interval may merit options, shared inventory or more frequent review. Low confidence does not mean no opportunity; it means the plan should preserve the ability to change course.

Watchlists connect analysis to action. Broader state contribution, improving fuel diversity and repeated outcomes above the central band can strengthen an upside view. Rising dependency, negative revisions and consecutive lower-band misses weaken it. A source-definition change or structural shock can invalidate the model entirely.

A forecast field succeeds when it improves decisions before the outcome is known. Its quality lies in coherent numbers, calibrated uncertainty, transparent revision and explicit response rules. The point estimate remains useful, but it is no longer allowed to impersonate the future.

The same discipline improves retrospective learning. Score each horizon, measure interval coverage and compare revisions with the signals named in the prior edition. A forecast that misses for an anticipated reason may call for recalibration; one that misses because the system boundary changed may call for a new model.

Over time, this record is more valuable than a sequence of isolated predictions. It shows where uncertainty is reducible, where it is structural and which signals repeatedly precede useful revisions. Forecasting becomes an organisational learning process, not a monthly contest over one number.

• Budget against the base, but stress-test the 80% range.
• Reserve irreversible action for signals that persist across releases.
• Use the 95% range for resilience and liquidity questions, not routine targets.
• Record forecast revisions as changed beliefs and explain the evidence behind each change.