Demand Forecasting

Demand-forecasting agents treat conversational, qualitative forecast adjustments (“bump it up for the campaign”) as replacements for structured baseline computations rather than deltas, embedding-retrieve cold-start demand analogs who are topically similar but demand-anomaly-laden, ignore structural constraints (promotion cancellations, bullwhip amplification) visible in multi-agent planning notes, and mismodel promotional lift and new-product adoption using flat, uninformed multipliers. These patterns cluster around two categories: structural errors (forgetting to apply conversational adjustments to the model, ignoring multi-agent coordination constraints) and data-driven errors (analogs and seasonality selected without domain grounding, lift multipliers not category-specific). Demand-forecasting errors propagate directly into inventory and production commitments; a 10-20% forecast error causes excess or shortage with full operational impact.

Key Takeaways

• 7 distinct failure patterns affect demand forecasting, grouped into four mechanisms: conversational-adjustment architecture (discarding baseline structure), cold-start analog selection (topical similarity vs. demand-driver similarity), constraint visibility across multi-agent workflows, and lift/adoption modeling (flat multipliers vs. category-specific elasticity).
• Bullwhip-effect and seasonal-demand failures are documented at “very common” frequency, with variance amplification of 5x typical at manufacturing tier when demand visibility is limited to downstream order data.
• Promotional-lift overestimation affects 15-30% of promotional-period forecasts when using flat historical multipliers without accounting for pull-forward and cannibalization, producing post-promotion overstock and markdown losses.
• Cold-start forecasts for new products show 30-50% higher error than established-SKU forecasts in the first 4-8 weeks when using a single historical analog or generic ramp curve instead of category/channel-segmented templates.

Scope

• Conversational Adjustment Architecture — conversational-forecast-adjustment-discards-structured-model-baseline. When planners request forecast adjustments conversationally, agents regenerate new numbers via free-text reasoning instead of applying a bounded delta to the baseline’s structured components (seasonality, trend, base rate).
• Analog Selection & Anomaly Blindness — embedding-retrieval-pulls-discontinued-sku-as-demand-analog-for-new-product, new-product-cold-start-misforecast. Cold-start demand estimation selects topically similar historical SKUs without screening for demand anomalies or uses a single global ramp curve instead of category/channel-segmented curves.
• Multi-Agent Constraint Loss — multi-agent-handoff-drops-promotion-cancellation-update-before-demand-forecast-run. Promotion-planning agent cancels a promotion recorded in free-text notes but does not update the structured promotional calendar the forecasting agent reads, so forecast still includes cancelled-promotion lift.
• Structural Demand Dynamics — bullwhip-effect, seasonal-demand-misses. Small downstream demand variations cascade upstream; seasonal patterns not learned from insufficient historical data; lead-time variance not modeled.
• Lift & Elasticity Modeling — promotion-lift-overestimation. Flat historical lift multipliers applied universally do not account for category-specific elasticity, pull-forward, cannibalization, or discount depth.

When Demand Forecasting Matters

• Demand forecasts directly drive production, procurement, and inventory commitments with long lead times; a 10% forecast error requires 10% excess or shortage, with full capital and markdown impact.
• Fast-moving domains (retail, e-commerce, CPG) have demand shifts that outpace quarterly planning cycles; stale historical analogs or flat multipliers can miss category-specific or promotion-specific dynamics entirely.
• Multi-agent supply-chain workflows introduce coordination points (promotion planning, production scheduling) where constraint visibility is low; a promotion cancellation invisible to the forecasting agent cascades into overproduction and waste.

Cross-Pattern Insight

All seven demand-forecasting patterns share a root cause: the agent prioritizes generative plausibility over structural fidelity. When adjusting a forecast conversationally, generating a smooth new number feels more responsive than applying a transparent delta to a complex decomposition (base rate, seasonal index, trend). When selecting a cold-start analog, embedding similarity to topical description (new product is a “wireless earbud”) is easier than structured matching on demand-driver type (is it a staple or a trend-driven item?). When modeling promotional lift, a single historical multiplier (3x sales) is simpler than category-specific elasticity tables. When forecasting seasonal demand, a global average ramp curve is easier to deploy than category/channel-segmented curves. Mitigation requires architectural constraints: mandatory delta-application for conversational adjustments (never free-text regeneration), structured-attribute pre-filtering for analog selection, shared event logs between planning and forecasting agents, and category-specific elasticity and adoption models with explicitly-acknowledged assumptions.

Frequently Asked Questions

How do you distinguish between a conversational adjustment and a forecast-model update?

A conversational adjustment (“bump it up for the campaign”) should be routed through the forecasting model’s own adjustment/override interface with a structured delta parameter. A free-text agent reasoning (generating a new number based on “the old forecast plus the campaign effect”) is not a legitimate adjustment path. Test: can you reproduce the adjusted number by applying the stated delta to the baseline? If not, the adjustment path is ungrounded.

What makes a cold-start demand analog “appropriate”?

An appropriate analog matches the new product on demand-driver type (is it a steady staple, seasonal, promotional, or trend-driven?), not just category or description similarity. Demand behavior of a mismatched analog carries over directly; using a trend-driven analog for a staple product, or vice versa, produces systematically biased forecasts. Pre-filter candidates by structured demand-driver classification before ranking by description similarity.

How do promotion cancellations get lost in multi-agent handoffs?

Promotion-planning agents record cancellations in free-text decision logs but demand-forecasting agents read from a structured promotional calendar. Unless the cancellation is immediately propagated to the calendar, forecasts generated after the cancellation decision still apply lift for a promotion that will not happen. Fix: require promotion-cancellation decisions to immediately update the structured calendar and trigger automatic forecast re-runs for affected windows.

What causes flat promotional lift multipliers to fail?

Promotional elasticity is highly specific to product category, discount depth, and whether the promotion is advertised. A flat 3x multiplier applied to all categories and depths misses low-elasticity categories (staples, necessity items) and overstates lift in high-elasticity categories. Flat multipliers also fail to account for pull-forward (demand borrowed from future weeks) and cannibalization (shift from full-price SKU to promoted SKU). Maintain elasticity estimates segmented by category, discount depth, and promotion type.

How do you test whether a seasonal model is catching real patterns or overfitting to training data?

Compare forecast error for seasonal periods against non-seasonal periods on a holdout test set. If seasonal error is not materially lower than non-seasonal, the model may be overfitting to training patterns that don’t generalize. Test separately on different seasonal years to check whether the learned pattern holds across years or is specific to the training window.

Patterns

Total: 7 patterns

Related Goals

• Inventory Optimization — downstream from demand forecasting; forecast errors directly drive inventory levels and safety-stock miscalibration.
• Logistics Routing — demand forecast drives shipment volume and timing; forecast error propagates into carrier-capacity and routing-resource constraints.