The program's own definitions (Pre-Read slide 8): short term is the incremental sales impact in the immediate weeks after an investment (W0 to W4), captured by media variables; long term is the impact that continues after that (W5+), captured by brand-health metrics - "mostly how brand equity is getting built up". The split matters because of two heuristics the trainers repeat in every session: all media together rarely explains more than ~12-20% of volume in the short-term read, and for a well-established brand the time-varying intercept sits at roughly 45-55% of volume (it can run anywhere from ~10% for a young brand to ~60% for a category stalwart).

That intercept is not noise. It is where brand equity, category effects, and everything the ST model cannot name accumulate. If you stopped at the short-term model, half the business would sit in an unexplained bucket and upper-funnel channels like TV - whose whole job is to build that bucket - would look like poor investments. The long-term construct exists to open the intercept up and hand its volume back to the media that earned it. Module 2.3's state-space machinery is what makes the intercept time-varying in the first place; this module is about what happens to it next.

MDS is the program's shorthand for the three pillars of Kantar's proprietary brand-health tracker (the bgs feed on the Data sheet, referred to in sessions as the BTS / "branded" data):

Mechanically: the pillars arrive as indexed values with no fixed range, where ~100 is the category average - observed values for strong brands commonly run well above that. The tracker is monthly (sometimes weekly); monthly values are replicated to each week within the month, since brand health does not move week to week. The leading pillar can switch over time - it has happened in live markets. A quick classification read: average each pillar over the window and see which leads; an established, non-innovating brand will usually come out salience-heavy.

The long-term construct is a relay of three model layers. All numbers below are synthetic, chosen to mirror the shape of the trainers' worked example for a well-established category leader.

STL1: the short-term model produces the LT dependent

Take a synthetic brand selling 20M units a year. The ST model attributes media, price, promo, distribution and macro, and leaves ~54% - about 10.8M units - in the time-varying intercept plus seasonality. For a stable established brand that intercept barely moves week to week (a few points around its level); a COVID-scale shock is what a big drop looks like. Two things the analyst does not control: the intercept and seasonality coefficients. You bound the drivers; the decomposition machinery owns the rest.

LTL0: splitting the intercept into brand equity

That 10.8M of intercept-plus-seasonality volume becomes the dependent variable of the next model, LTL0. It is regressed on the raw, untransformed MDS indexes - no adstock, no Hill, the index values as Kantar publishes them - plus a constant and an error term:

The one thing you do fix is the constant: ~25% of the intercept volume for an established brand - here 25% of 10.8M = 2.7M units. The trainers call this "true brand strength": consumer behaviour, category effects, everything that is "absolutely not explained, going out of the modelling purview altogether". It leaves the ecosystem; no channel ever gets credit for it. The engine then allocates the remaining ~8.1M across the three pillars in line with the index magnitudes - if the synthetic brand tracks at Meaningful 185, Differentiation 160, Salience 115, meaningfulness takes the largest share and salience the smallest. A weekly error term is inherent here: sales fluctuate weekly while MDS is flat within each month.

LTL1: three models running in parallel

Now each pillar becomes a dependent variable in its own right - three separate linear models, not one multi-output model - each with transformed media drivers, a constant, and an error term. The driver sets differ by pillar logic:

Distribution is the only base variable permitted anywhere in the long-term model, and only in the salience pillar (Methodology workbook rule). The three-model parallelism then continues all the way down the hierarchy: LT L2 (media level), L3 (platform), L4 (campaign) each run as three models - one per pillar - exactly as Pre-Read slide 13 draws it. This is not just a diagram: the Italy CIF long-term output folder contains three physically separate L3 files - LT_L3_differentiation_Signconstrained_italy_cif.xlsx, a ..._meaningful_... and a ..._salience_... twin - with identical sheet and column schemas. Three parallel runs, on disk, per level.

The bridge between the two horizons is a program-coined term: the Long-Term Conversion Factor. It is defined by the ST/LT split of a channel's total contribution:

Worked synthetic example: a developed-market TV channel carries an ST ROI of 0.8. TV's split heuristic is ~40/60, so LTCF ≈ 1.5. LT ROI = 0.8 × 1.5 = 1.2, and the total ST+LT ROI = 0.8 + 1.2 = 2.0 - which is exactly the trainers' rule of thumb that a leader's TV lands "around two-ish" all-in, from an ST number that on its own would look like a losing bet. This is the whole commercial point of the LT construct.

The channel bands, as the trainers stated them in the KT sessions (heuristics, not statistical outputs - they encode funnel position and consumer behaviour):

Developed vs developing markets skew everything here. In a developed market (UK, France) expect ~0.6-0.8 back short term per unit invested and "structured" total ROIs - do not expect 4s and 5s. Developing markets (India, Brazil) can plausibly show ROIs of 4-5, and LTCFs run higher too: consumers are more receptive and adaptable, and the category has headroom. The same channel, same spend discipline, benchmarks completely differently across the two - which is why LTCF is set per market, never copied across.

A philosophy point from the KT(4) session worth carrying into every cell: the first run is always unconstrained. Factors can come back at absurd magnitudes - the trainers' own words allow for "-500%, +2000%" - but that run is not a result, it is a compass: it shows the direction and the relative magnitude the data wants to give each driver. A factor that correlates very strongly with the KPI might then justifiably be bounded at 40-45% where the default heuristic said 30%. Only after that read do the constraints go on - a sign constraint and a magnitude constraint per variable, with the invariant that everything must add back up to 100% of volume. Bounds are a dialogue with the model, not a cage bolted on before the model has spoken.