Primary research turns intent into an explicit prior

When an apparel venture lacks historical transaction records across price changes, managers frequently ask primary research to deliver a single price-elasticity coefficient (for example, claiming that a 10 percent price cut will raise demand by 18 percent across the market).[S03][S31] Treating a survey-derived price slope as a universal structural elasticity is a category error.[S48][S0502] Choice experiments measure willingness-to-pay trade-offs conditional on the competing profiles shown on the interview card, the respondent's immediate framing and the sampled channel; they do not observe general-equilibrium competitor retaliation, credit constraints among wholesale buyers, or seasonal liquidity shocks.[S06][S31][S48]

In Cameroon's clothing market, price response is highly non-linear and exhibits sharp threshold effects around psychological currency notes (such as XAF 2,000, XAF 5,000 and XAF 10,000) and existing channel reference prices.[R14][R26][R32] Against an average per-capita clothing and footwear expenditure of about US$94.70 a year (roughly XAF 57,500 a year at illustrative reference exchange rates) across roughly 8.0 pieces per person, an everyday garment priced near XAF 3,500 to XAF 5,000 competes in a mass replacement tier, whereas a garment priced above XAF 12,000 shifts into an occasional or formal discretionary tier.[R26][R32] A conjoint model fitted across three discrete price levels (for instance, XAF 5,000, XAF 6,500 and XAF 8,000) identifies the local utility difference between those tested nodes for the sampled respondents.[S31][S48] Extrapolating a linear log-price coefficient outside that tested interval, or applying a boutique shopper's price slope to an open-air market stallholder, fabricates precision where no evidence exists.[S08][S0502]

Consequently, GFT forbids encoding primary price research as a continuous elasticity parameter inside the automated forecasting service.[S33][S36] Instead, the discrete price tiers tested in the conjoint instrument are stored as explicit conditional scenarios (Low-Price Penetration Scenario, Baseline Margin Scenario, Premium Positioning Scenario), each carrying its own full quantile distribution ($q_{0.05}$, $q_{0.25}$, $q_{0.50}$, $q_{0.75}$, $q_{0.95}$) and sample provenance tag.[S08][S16]

A second hazard in pre-launch price analysis is ignoring the interaction between unit price, stock availability and substitution when the first pilot lots enter physical shops.[S06][S07] Suppose a venture tests two price cells in a twelve-week pilot across matched retail outlets, placing a basic cotton T-shirt at XAF 4,500 in Cell A and XAF 5,500 in Cell B.[S32] If the XAF 4,500 outlets experience stockouts in medium and large sizes during week four because replenishment was sized to the central prior rather than the 80th percentile, raw sales volumes in Cell A will be censored downward.[S06][S07] Comparing uncensored sales in Cell B against stockout-censored sales in Cell A would falsely imply that buyers are price-insensitive, when in reality latent demand at XAF 4,500 was truncated by empty shelves.[S07]

Therefore, no survey-based willingness-to-pay distribution is updated or replaced by pilot price-response estimates until the pilot transaction log has been audited for stockout timestamps, size-run completeness and concurrent promotional discounts as specified in Chapter 2 and Chapter 7.[S06][S07][S37] Until uncensored, randomized or quasi-experimental price variation is observed in live transactions, price remains a scenario selector owned by commercial leadership rather than an autonomous feature weight.[S32][S36]

By the close of pre-launch fieldwork, a greenfield venture holds three distinct streams of evidence: the analogue launch pool from Chapter 3, the macro and trade bounds from Chapter 4, and the calibrated choice surveys and expert Delphi distributions from the present chapter.[S08][S28][S31] Combining diverse forecasts is one of the most reliable ways to reduce variance and improve accuracy in forecasting practice, with over fifty years of empirical research showing that simple, transparent weighting schemes such as equal-weight or pre-fixed precision combinations consistently rival or beat complex optimised weights.[S20] However, forecast combination delivers its variance-reduction benefit only when the constituent forecasts contain independent information rather than echoes of the same underlying belief.[S08][S20]

In a small industrial project, information contamination across evidence streams happens easily and silently unless it is policed.[S20][S37] Consider three common failure modes:

To prevent double counting and post-hoc weight shopping, GFT requires every pre-launch evidence stream to be logged in a formal Prior Dependency Register before the individual readouts are combined, as illustrated in Exhibit 25.[S20][S36] For each stream, the register records the responsible owner role, the exact sample size or reference-class basis, any shared data or informational overlap with other streams, and the pre-fixed combination weight summing to 1.00 across eligible sources.[S20][S36] In the illustrative configuration of Exhibit 25, the analogue launch pool and the calibrated buyer conjoint survey each carry an equal weight of 0.35 because they rest on distinct empirical bases (historical launch curves versus prospective buyer choice calibration), while the wholesale structured-analogy interviews and the anonymous Delphi panel each carry 0.15 because both rely on practitioner memory and overlap partially with three historical lines in the analogue pool.[S20][S28][S31] While Exhibit 25 makes assumptions and weights transparent for audit, the table itself does not prove that the chosen weights are optimal; their empirical skill can be scored only after pilot actuals arrive.[S14][S20]

When combining probabilistic priors from multiple evidence streams, the mathematical aggregation rule matters as much as the weights.[S14][S20] If an analyst averages the 10th percentiles, 50th percentiles and 90th percentiles across sources that disagree on the central level, or uses a naive precision-weighted normal conjugate update that assumes independent Gaussian errors, the resulting combined distribution becomes artificially narrow precisely when the sources strongly conflict.[S14][S28] Instead, GFT combines pre-launch distributions via a finite mixture (linear opinion pool) of the component predictive densities, $F_{\text{prior}}(y) = \sum_{m=1}^{M} w_m \, F_m(y)$, where $\sum_{m=1}^{M} w_m = 1$ and $w_m \ge 0$.[S14][S20]

Under linear opinion pooling, if the analogue launch pool ($w_1 = 0.35$) centres around an illustrative median of 800 units and the calibrated buyer survey ($w_2 = 0.35$) centres around 1,018 units while the wholesale analogy panel ($w_3 = 0.15$) warns of a downside cluster near 450 units, the combined prior distribution retains the variance between the streams as well as the variance within each stream.[S14][S20] Disagreement among pre-launch sources widens the 80 percent and 95 percent predictive intervals rather than cancelling out into a false sense of certainty.[S14][S16]