01 · Establish the comparison

A similar watch needs a similar story.

Match franchise, price, CMF, region, distribution, and weeks since launch. A franchise collector release needs comparable collector launches. An ultra-limited drop can serve as a sustain campaign, keeping interest in the franchise after launch. Its supply-constrained sales should be compared with similar campaigns, not used as the demand curve for a volume launch.

Top-line baseline: (3,360 + 5,040) ÷ 2 = 4,200 units. The comps are alternatives for one launch, so average them; do not add them. Equal weights make the starting assumption transparent.

Clustering narrows the search. Merchandising checks the fit.

Across earlier launches, group products by franchise, price, CMF, planned channel reach, and launch support. Scale numeric features and encode categories using the training history. Use attributes known before the new launch; its later sales cannot influence which comps are selected.

The output is a shortlist with reasons for inclusion, not a forecast. Merchandising reviews the nearest relevant group and selects A and B. Keep collector releases and sustain drops separate from volume launches; if no credible group exists, document the weak match instead of forcing one. The equal-weight average remains the benchmark.

Method reference: An Introduction to Statistical Learning · Clustering.

02 · Find the normal selling pattern

Keep load-in out of the DTC pace.

Wholesale’s large pre-launch orders belong in the launch commitment, but not in the consumer selling-rate sample. Exclude those weeks and any stockout weeks using launch dates, order records, and availability flags. Missing demand during a stockout is not zero demand.

Apply TRIMMEAN to the eligible, clean weekly records to estimate the normal pace. In this example, A returns 280 units/week and B 420 units/week: an average of 350 units/week, or 2,450 DTC units over seven selling weeks, before launch adjustments.

This steady-flow rate is a DTC proxy until channel records validate it. TRIMMEAN trims both tails; it does not identify wholesale orders, detect stockouts, or assign channels by itself.

Keep the actuals. Preserve the correction.

Retain original weekly sales, returns, channel, product age, campaign exposure, and availability records. Alongside them, store imputed historical demand: an estimate of what could have sold during a confirmed stockout. Keep observed units, estimated lost units, and the combined history in separate fields.

Week 4’s actual sales remain 0. The 280 estimated lost units bring its adjusted history to 280. Use that adjusted history when comping the next NPI so limited availability is not carried forward as weak demand. For partial stockouts or a changing launch curve, refine the estimate using in-stock exposure and matched product-age weeks.

Uncorrected history: 3,850 × 0.75 × 1.20 = 3,465 units
Stockout-adjusted history: 4,200 × 0.75 × 1.20 = 3,780 units

The correction adds 315 units to C’s illustrative plan. That is the effect of the demand estimate, not proof those sales were lost or that the next launch will succeed.

03 · Normalize the launch conditions

Adjust for C’s marketing support.

A and B received paid ads, a site hero, and email. C gets email only. Use an illustrative 0.75 marketing factor, then a 1.20 seasonal factor for C’s stronger selling period. A weaker season would use a factor below 1.00.

4,200 × 0.75 × 1.20 = 3,780 units

These factors are assumptions, not measured marketing lift. Estimate them from comparable exposure or controlled tests, keeping season, price, and availability separate. If the marketing story changes, revise the assumption and save a new forecast version.

Regression estimates the adjustment.

Fit weekly demand across earlier launches using product age, calendar season, price, channel reach, and actual marketing exposure. Product age and calendar time are separate: a launch-week spike should not become a holiday effect. Keep observed sales and imputed demand flagged, and check whether conclusions change when estimated weeks are excluded.

Predict the same product under the comps’ launch support and under C’s email-only plan in its target season. That comparison can replace the assumed 0.75 and 1.20 factors. Model the effects together and allow channel responses to differ; do not apply the same marketing or seasonal uplift a second time. If hero placement, ads, and email always occur together, the history cannot reliably separate their individual effects.

Method reference: Forecasting: Principles and Practice · Marketing and seasonal predictors.

Experiments test what marketing actually adds.

Regression can predict an association; it does not establish incremental sales. Where feasible, randomly assign eligible audiences to email-only support or the added campaign. Agree the test window and success measure in advance, hold price and product availability comparable, and compare units and contribution per assigned customer. Check total purchases across channels so a channel shift is not mistaken for new demand.

Use the estimated difference and its uncertainty to inform the next marketing assumption. Check that the tested audience and campaign resemble the next launch before transferring the result. If randomization is impractical, retain the historical relationship as a planning estimate rather than label it causal lift.

Method reference: Microsoft Research · Controlled experimentation.

04 · Reconcile the channel plan

One total. Two different demand signals.

The 1,575-unit wholesale balance is an allocation hypothesis, not proof of consumer demand or a firm order. Reconcile it against account orders, opening inventory, return rights, and retailer sell-through. One buyer’s load-in is a different signal from many individual DTC purchases.

For simplicity, both channels receive the same adjustment factor. Real channel responses can differ. This example covers DTC sales and brand shipments to wholesale; retailer sales are a validation signal, never added again. Include marketplace as its own channel where present.

Decision trees make segment differences discussable.

Fit a small, pruned tree on earlier launches to test whether attributes such as collector status, price band, or distribution reach explain DTC share or channel forecast error. Require enough launches in each branch and stable performance on later launches. A rule such as “collector release with broad DTC reach” is a candidate to test, not a finding asserted by this example.

The output is an understandable segment rule for reviewing allocation. Use a separate rule only if it improves on the pooled plan; otherwise retain the common baseline. Reconcile any revised DTC share to the same total and validate the wholesale remainder with account evidence.

Method reference: An Introduction to Statistical Learning · Decision trees.

05 · Learn as the launch unfolds

Update the remaining curve, not the past.

The adjusted comp curve supplies the opening forecast. As each week closes, compare clean, in-stock DTC demand with the corresponding week since launch. Test a time-series update that smooths the observed-to-planned demand ratio and, if supported by enough history, its trend. Learn the update strength on older launches so one unusually strong week does not rewrite the entire season.

Project that update over the remaining launch-age curve, using only campaign plans and evidence available at that review. Model wholesale load-in and replenishment separately; a bulk order should not reshape the DTC curve. Treat stockouts as constrained observations and report their estimates separately.

Decision output: a new forecast version by channel and week, with a plausible demand range for Supply to review. Preserve the original launch commitment. Compare an updated curve with the unchanged comp curve at the same review date and over the same remaining horizon.

Method reference: Forecasting: Principles and Practice · Updating level and trend. The ratio update described here is a candidate adaptation to launch demand, to be tested against the simpler curve.

07 · Turn the evidence into a commitment

Update the next decision, not the old forecast.

Demand planning owns the baseline and exceptions. Marketing confirms exposure; merchandising validates the comps; Sales reconciles wholesale orders and returns. Supply checks the next available delivery, and Finance reviews the cash and inventory exposure.

At each review, forecast by channel through manufacturing lead time plus the fastest feasible transit and receiving time. Net against usable inventory and confirmed inbound supply; set any safety stock from forecast error and the service target. The launch forecast is an input to that decision, not an automatic purchase order.

After each launch, retain its observed weekly sales and publish a separately versioned, stockout-adjusted history for future NPI comparisons. Keep forecast versions and override reasons. Evaluate service, excess inventory, returns, and expedites alongside forecast error. This illustration demonstrates the method; it does not establish a measured business improvement.