Every retailer in 2026 is shopping new demand forecasting models. The vendor pitches are all variations of the same theme: foundation models, time-series transformers, agentic forecasting, multi-modal signal fusion. The CDO is taking three to five vendor meetings a quarter trying to figure out which model to bet on.
The diagnosis is wrong. Demand forecasting at most retailers in 2026 doesn’t have a model problem. It has a signal problem.
Replacing the model without fixing the signal layer is a 12-18 month project that produces a 2-5% accuracy improvement and consumes the data team’s calendar. Fixing the signal layer first produces 8-15% accuracy improvement on the existing model, and unlocks the new model’s actual value when it eventually gets deployed.
Most retailers are doing this in the wrong order.
What “signal problem” actually means
Demand forecasting is a function: signal in → forecast out. The model is the function. The signal is the input.
When the forecast is wrong, retailers default to “we need a better function.” That’s the model conversation. But the function is usually doing what it’s supposed to do — given the input, the output is roughly the model’s best estimate. The wrongness comes from the input, not the function.
Three signal-layer failures I see at scale:
Failure 1: Promotional uplift signal is decoupled from the forecast
The retailer runs a promotional campaign. The forecast doesn’t know. The forecast predicts baseline demand; the actual demand is baseline + promotional uplift; the forecast looks “wrong” by the size of the uplift.
The fix is not a better forecasting model. The fix is wiring the trade-promo system into the forecast as a signal — what’s promoted, when, at what price, in which retailers, at what scan-back rate. Same model, much better forecast.
Most retailers in 2026 still have these systems disconnected at the data layer. The merchandising team’s promo plan lives in one system; the forecasting team reads SKU-level baseline demand from another. The integration debt is the bottleneck.
Failure 2: Inventory state signal is stale
The forecasting system reads yesterday’s inventory state. Demand on Tuesday is shaped by what’s actually on the shelf Monday night. If Monday-night state is overrepresented in the forecast (e.g., reflecting a stockout that the system hasn’t yet captured), the forecast for Tuesday is wrong by the size of the unobserved real-time state.
Fix: real-time inventory state, not yesterday’s. RFID, IoT, or POS-derived. The forecasting model doesn’t change; the signal layer becomes current. Forecast accuracy improvements show up immediately.
This is the RFID-coming-back-from-the-dead story. Real-time inventory state is the missing input most forecasting deployments need; agentic operations make it justifiable to bring RFID back online for inventory observability.
Failure 3: External signal is treated as decoration, not input
Weather. Local events. Competitor stockouts. School schedules. Holiday-shifted shopping behavior. These are real demand-shaping signals. Most retailers either ignore them or include them as a checkbox on a vendor’s “additional features” tab without actually wiring them into the forecast as first-class inputs.
The forecasting model’s accuracy ceiling is bounded by what it knows. If it doesn’t know that a local high school’s homecoming game shifted the weekend demand pattern in 3 stores, the forecast for those 3 stores is going to be wrong, and no model upgrade fixes that.
The diagnostic: model problem or signal problem?
Three questions to surface where the failure actually lives. Run them before any vendor meeting.
1. What does the forecast residual decompose into? A residual analysis of the forecast errors — by SKU, by store, by day, by event — should show systematic biases. If errors cluster around promotional events, that’s a promo-signal issue. Around weather events, weather-signal. Around stockouts, inventory-signal. Around the same 5% of stores consistently, store-attribute signal. Around random noise across the whole portfolio, that’s a model problem (and usually the smallest of the four).
If your team can’t decompose residuals at this granularity, the forecast quality conversation is happening at the wrong abstraction layer. Vendor pitches at that abstraction level are designed to obscure the real problem.
2. What signals does the model actually consume — and are they fresh? List the input fields the model reads. Look at when each was last updated. Anything refreshed slower than the forecast cadence is a candidate for staleness-driven error. Forecasts run daily but inventory updates nightly? That’s a 24-hour gap. Forecasts run weekly but promo plans update intra-week? That’s missed signal.
A fresh-but-not-immediate input is often correctable without a model swap — just by raising the input’s refresh cadence.
3. What signals does the model NOT consume that you have available? Inventory of the signals available in your stack but not in the model: third-party weather, retailer scan data, intra-day POS, local-event calendars, competitor-pricing scrapes, seasonal merch event calendars, customer-loyalty redemption velocity. For each one, ask: “would including this materially improve the forecast?” Most retailers have 5-10 unconsumed signals that would; the model is currently flying blind on them.
The retrofit playbook
A retailer with these three failures has 12 months of work that doesn’t involve replacing the forecasting model:
Quarter 1: signal observability. Instrument every input the current model consumes. Track freshness, completeness, and contribution to forecast residual. Identify the top 3 signals that, when fixed or added, would close the largest residual category.
Quarter 2: signal layer build. Wire the top 3 signals into the model. Trade-promo integration, real-time inventory feed, external signal ingestion (whichever apply). The forecasting model itself doesn’t change in this phase — only its inputs.
Quarter 3: re-evaluate the model, with cleaner signal. Now ask “is the model the bottleneck?” Most retailers find that residual error has dropped 8-15% from signal improvements alone, and the remaining error is roughly evenly distributed (not systematically biased) — which is the case where a model upgrade actually helps.
Quarter 4: model deployment, if warranted. With clean signal, a new forecasting model’s improvement shows up in measurable accuracy gains rather than getting absorbed into the same signal-noise floor.
The retailers running this sequence in 2026 are pulling away from the cohort that’s still in vendor meetings.
Why the model conversation persists despite the signal evidence
Three reasons the model-first framing wins inside retailers, even when the data points to signal:
1. Vendor pull. Forecasting vendors sell models. Their pitch is naturally model-shaped. They don’t have a “let us help you fix your signal layer” SKU.
2. Internal politics. “We need a better model” is a vendor-purchase conversation. “We need to fix our signal layer” is an internal-engineering conversation that involves the data team, the trade-promo team, and the merchandising team negotiating data integrations. The first conversation is easier to convene; the second is the one that actually moves the needle.
3. The data team owns the model; the operations team owns the signal layer. When the forecast is wrong, the data team gets the question. They look at the model. The model is doing what it’s supposed to. They escalate as “we need a better model.” The signal-layer conversation never happens because the data team can’t unilaterally fix it — the inputs come from teams the data team doesn’t run.
The CDO’s job is to convene the cross-team signal conversation. Without that, the signal layer drifts further out of repair quarter over quarter, and the model upgrades get progressively less effective.
The CODN angle
The cost of doing the model upgrade first instead of the signal fix:
- 18 months of vendor evaluation, model selection, and migration consuming the data team’s senior capacity, with a 2-5% accuracy improvement to show for it.
- Same systematic forecast biases that drove the original “model problem” framing — because the signals haven’t changed.
- Talent flight risk as the senior data engineers who can do signal work get poached by retailers actually doing it.
- Optionality decay — the signal layer work is the prerequisite for every adjacent forecasting deployment (promo optimization, demand-driven replenishment, agentic merchandising). Skipping it doesn’t just delay forecasting; it delays everything downstream.
The CODN of model-first sequencing at Tier 1 retail scale is conservatively 12-18 months of compounding capability loss against the cohort that fixed signal first.
The bottom line
Demand forecasting in 2026 has a signal problem, not a model problem.
Decompose the residual. Find the systematic biases. Fix the signal layer. Most of what looks like “model failure” is signal failure wearing a model costume — and most of the vendor pitches in your inbox right now are selling you the wrong fix.
The retailers running the signal-first playbook in 2026 are going to look unrecognizable from the ones still shopping forecasting vendors in 2027. The data is in the residuals. Read it.