Bound setting is the program's proprietary craft: before the constrained model search runs, every variable receives an allowable coefficient range (LB/UB). The convention: base variables are bounded on contribution; media and other incremental factors are bounded on ROI (module 1.7). Two disciplines frame everything:

• Everything must add back to 100%. The % volume contribution column across all drivers, intercept included, totals 100. Size the intercept first - ~45-50% for a very well-established brand - and you immediately know how little room remains for everything else. The intercept is high because brand equity, category effects and everything unmodelled live there (it later feeds the MDS split, Track 2.4).
• Let the model speak inside the bounds. The first run at every level is unconstrained apart from signs - factors can swing to absurd magnitudes, but that run gives direction and relative size. Bounds then give the model a corridor, not a script. Reverse-engineering bounds from the due-to targets you want is a named pitfall: it takes the freedom out of the model.

Order of operations: intercept, then the base ladder below starting from the biggest driver per the EDA (usually distribution for an established CPG brand), then media, then competitor and halo, macro last and small.

Each band comes with its condition and its reason. The bands are method, not client data - they travel across engagements.

Two cross-checks close the base side. The promo-intensity efficiency check: promo intensity = promo sales volume over own sales volume; total promo contribution (promo price + promo TDP together) should land at roughly 50% efficiency against it - a brand with 40% promo intensity should show total promo contribution near 18-20%, never the full 40. And the base-vs-incremental top-down check: when total media ASSR is around 4%, base-driven contribution should be at least ~80-85% - a mature brand's investments maintain awareness, they do not create the business.

The EDA correlations (module 1.4) feed the bands through a simple mapping in the Target bounds theory sheet:

Use correlation with judgement, not as a formula input. Its best use in KT practice is dispersion: set the national/total-sub-brand bound as the median, then push regions above or below it by their correlations. Across different sub-brands, correlation transfer is invalid - investment strategies and sales shares differ. And two correlations are explicitly untrustworthy: aggregated competitor price and aggregated competitor spends (spurious by construction; use the category snapshot instead).

The levy and the final widths

Once a contribution is set, the bound is not a point - it gets a plus-minus 15% levy around the set value (so a -7% price contribution becomes a band of roughly -6.5% to -8%). At the final model stage the widths settle at base variables plus-minus 20% and media plus-minus 5 to 7.5%: base gets more room because the EDA evidence behind it is strong and its share is large; media is held tighter because its expectations are benchmark-anchored to ASSR and small errors there distort ROI stories directly.

Targets become config-ready coefficient bounds through a small factor method, one row per variable:

Reading it: current is what the last model run delivered for this variable, target is what the ladder says it should deliver; their ratio rescales the current coefficient toward the target, and the levy opens a corridor around it. A CONCAT column then assembles the literal string pasted into the config - contrib_<variable>: [LB, UB] - flipping LB and UB order when the coefficient is negative so the smaller number always leads. The long-term block repeats the identical pattern three times, once per MDS pillar. This loop (model run → compare to targets → recompute factors → new bounds → re-run) is the Validations-Modeling handoff cycle in the Process SOP, and it continues until targets are met.