In a budget review, when the variance between prior-period actuals and forecast comes up for discussion, the first question asked is almost invariably the same one: which line item accounts for the gap. The item is identified, its circumstances are rehearsed, the next period's forecast is adjusted on that basis, and the meeting concludes. The following quarter produces another variance, another line item is located, another adjustment is made. Even when the same sequence repeats a third time, the discussion remains anchored to individual items; the fact that every variance has fallen on the same side of the estimate — that the model is drifting systematically in one direction — rarely reaches the agenda as a subject in its own right.
The same pattern recurs on the commercial side. Conversion rates, sales cycle length, customer acquisition cost: once calibrated, these become planning assumptions, and headcount, marketing budget, and working capital requirements are all derived from them. The coefficients are drawn from trailing twelve- or twenty-four-month averages, and precisely for that reason they dilute whatever information the most recent period carries. When market structure shifts, when competitive intensity changes, or when purchasing authority migrates within the buyer organisation, the model registers the change only after several periods of lag, and then only in proportion to the weight the average assigns to recent observations.
The mechanism beneath this picture is termed nonstationarity — the condition in which the statistical relationship between variables does not remain constant over time. Every forecasting structure assumes, implicitly or explicitly, that a relationship observed in the past will hold going forward; without that assumption, inference from historical data to future outcomes is not possible at all. The difficulty lies not in the assumption but in the absence of any monitoring of the conditions under which it remains valid. When a relationship breaks, the model does not return an error, does not halt, does not declare itself void; it accepts inputs, applies the formula, and produces output whose formal tidiness is entirely independent of its substantive validity.
What makes this mechanism unusually persistent in a corporate setting is that reliance on historical relationships is, most of the time, the correct behaviour. On a repeating production line, within a mature customer portfolio, in a market whose regulatory regime has settled, historical coefficients describe the coming period with reasonable accuracy, and rebuilding the justification for every forecast from first principles would raise decision costs to an indefensible level. The shortcut is rational to the extent that it lowers cost while conditions hold constant. Deterioration begins when the conditions change and the choice does not — and because the drift is typically incremental, the break itself never appears large enough in any single period to command attention on its own.
A second layer concerns the confidence attributed to the length of the data series. Intuition holds that a longer series yields a more reliable forecast; under nonstationarity, that relationship inverts. If the structure has broken, the longer window carries observations belonging to the prior regime into the current estimate with substantial weight, pulling the model toward conditions that no longer obtain. A three-year average will, in that situation, produce a tighter statistical confidence interval alongside lower forecast accuracy; the narrowness of the interval is a function of sample size rather than of correctness, and this distinction is routinely lost by the time the figure reaches a presentation slide.
The first surface on which the cost appears is working capital. Inventory policy is constructed from lead times and demand coefficients; when the demand structure shifts, the inventory line does not grow — turnover slows, which is to say the balance sheet shows not a conspicuous increase but the same inventory level carried for a longer period. Receivables behave comparably: as payment behaviour on the buyer side changes, days sales outstanding accumulates in increments of a few days at a time, lengthening the cash conversion cycle without ever announcing itself. When both items drift together, financing requirements expand quietly while the profitability statement shows no deterioration whatsoever.
The second surface is on the credit side. Covenant thresholds in a loan agreement are calibrated to the revenue and margin relationship prevailing at signing; when that relationship drifts, the thresholds remain fixed while the headroom against them does not. Here nonstationarity assumes its most expensive form, since the model's drift no longer stays contained as a planning error but converts into a contractual trigger, transferring the company's negotiating position outward at precisely the moment that position is weakest. The same logic governs earn-out structures: a post-closing performance threshold written with pre-closing coefficients moves, once the regime shifts, in a direction neither party priced at signature.
The third surface is valuation. In a diligence process, what receives scrutiny is less the forecasting model itself than the model's track record of accuracy across prior periods; a model that has erred in the same direction over consecutive periods raises, on the buyer side, not merely a question of forecast quality but a question about management's capacity to read its own market. The consequence surfaces not directly in the multiple but in the structure of the transaction — broader representations and warranties, a higher escrow proportion, additional conditions precedent. The question a company rarely puts to itself is this one: what evidence demonstrates that the relationship generating this forecast remains in force.
The structural intervention is not more frequent recalibration; frequent recalibration accommodates the drift rather than correcting it, and obscures the moment of break altogether. A functioning architecture has four components. The first is that the assumptions producing the forecast be written into a separate record at the moment of proposal rather than at the moment of approval: which coefficient, drawn from which period, valid under which condition. The second is that the sign of the variance be tracked independently of its magnitude — a forecast erring in the same direction across three consecutive periods triggers structural review regardless of how small each error is. The third is that parallel forecasts be generated using windows of differing length, with the divergence between windows reported as an indicator in itself. The fourth is that the commercial relationships feeding the model be monitored at contract level, since regime change frequently appears in tenor, security, and pricing clauses before it appears in the data.
BEIREK operates this architecture on capital-intensive projects through an assumption register: every coefficient used in the financial model is held in a single record together with its source, its reference period, and the condition under which it holds, and that record lives as a document distinct from the model itself. When the model is updated, the subject of discussion is not the output but which assumption changed and on what grounds; recalibration thereby ceases to be a silent accommodation and becomes a recorded decision. Review cadence is tied to triggers rather than to the calendar — consecutive same-direction variance, a change in contractual terms on the supply or offtake side, a shift in the regulatory regime.
On the project finance side, the counterpart to that cadence is running covenant calculations not only for the current period but in parallel across scenarios in which the coefficients in the assumption register have moved; the sensitivity of headroom to each individual coefficient becomes known before the matter reaches a credit committee. The same drift means different things to different parties — for the sponsor it is a question about the timing of equity returns, for the developer a question about the closing timetable, for the senior lender a question about the scope of security — and the intervention works only to the extent that all three can consult the same assumption register.
The most dangerous property of a forecasting model is that it gives no notice when it has become wrong; the arithmetic continues to function, the output remains properly formatted, the quality of the presentation does not degrade. The operative question is therefore not how accurate the model has been, but on what evidence the relationship underpinning it can still be said to hold; and that question admits of an answer only where the assumptions are held in writing, in a place separate from the forecast they produce.