Module 2.1 established the hierarchy: an L1 model decomposes sales into channel contributions, then an L2 model decomposes each channel's contribution into platforms, then L3 splits platforms into campaigns. Each of those lower-level models is its own regression, and every regression leaves a residual. So when you sum the platform contributions that the L2 model produced, they do not add back to the channel contribution the L1 model handed down - the L2 residual is the gap. The same mismatch recurs at every split: platform → campaign, and (on the LT side) pillar-level models summed across Meaningful, Differentiation and Salience.

Left uncorrected, this breaks the deliverable. The client reads a waterfall where Digital Video drives some volume at the channel page, flips to the platform page, adds up the platforms, and gets a smaller number. Numbers that don't tie are the fastest way to lose a read-out. Error allocation is the program's fix: push the residual back down into the children so every level re-sums exactly to the level above.

Two ways exist to distribute a residual: equally (same absolute error added to every child) or by weight (bigger contributors absorb more). The program uses weighted allocation - the working assumption being that a platform responsible for 45% of the modelled contribution is also plausibly responsible for about 45% of what the model missed. Equal allocation would distort small platforms violently (a tiny campaign could double or go negative) while barely touching large ones.

The mechanics are a single scaling factor per parent, per period:

Because every child is scaled by the same multiplier, each child's share of the parent is preserved exactly - that is what makes the allocation "weighted". After multiplication, the children sum to the parent by construction.

Here is the methodology workbook's own worked example (authored teaching numbers, not client data). The L1 model says the Digital Video channel contributes 989,787,900. The L2 model's platform contributions sum to only 880,911,231, so the factor is 989,787,900 ÷ 880,911,231 = 1.1236 (1.1235955 unrounded):

Note that the post-allocation column re-sums exactly to the channel's 989,787,900. Youtube, the largest platform, absorbed the largest absolute share of the residual (about 48.9M of the 108.9M gap); Hotstar, the smallest, absorbed the least (about 9.8M) - but every platform's percentage share of the channel is unchanged.

One rule specific to the long-term side: for LT error, first sum the contributions across the three pillar models (Meaningful + Differentiation + Salience), and only then compute and apply the factor. The pillars are three parallel regressions (module 2.4), so allocating per pillar before summing would triple-count the correction.

The sibling operation to error allocation is total-level aggregation: rolling results up from sub-brands, zones, or regional clusters to a unified brand view, so the client can read both the detailed and the total picture consistently. The trap is that not every quantity aggregates the same way. Contributions are volumes, so they always add. Spend adds. But rates and indices - a price, a distribution measure, a CPI - are not volumes: summing three regions' prices produces a meaningless number. The workbook's rule table:

Read the column split carefully: the contribution of price (its volume impact) sums across sub-brands like any other volume, but the price itself - the driver value you would display beside it - averages. Getting this wrong is invisible in the contribution waterfall and very visible the moment a client asks "so what was the average price?"

Synthetic numbers, fabricated for this exercise. The L1 model gives the Paid Social channel a contribution of 240,000 units. The L2 model produces these raw platform contributions:

Compute the factor, then each platform's post-allocation contribution, before opening the quiz. Check yourself: the three answers must sum to exactly 240,000, and each platform's share must still be 48 / 30 / 22.