Demand forecasting

Demand forecasting that starts with the best model for every pattern.

HZN-1 matches every product, customer and location pattern to its own best combination of models and publishes it as your baseline. Deep Sensing adds what the history misses, and only the key calls reach your planners.

Baseline forecast
SKU-4471 · Retailer North
HZN-1 live
Backtesting 16+ models and combinations on this item’s historyScored on accuracy, bias and stability Most accurate for this item: a blend of 3 modelsPublished as the baseline, retrained as new data arrives Today
Demand historyCandidate modelsHZN-1 baseline
Chosen by backtestWeight
Launches · causal45%
Gradient boosting30%
Proprietary neural networks25%
Output
The baselineThe scenariosThe drivers behind every number
Same data, same answer · retrained every week
18–45%

lower forecast error than traditional systems

Forecast error (illustrative)
Traditional system
HZN-1
16+models

compete for every item

Best blend for this item: 3 models
Weekly

retraining, or more often

W10W11W12W13
HZN-1best combination wins Machine learningDeep learningCausal modelsStatistical modelsFoundation modelsGradient boostingProprietary neural networksPrices, events and launchesThe initial benchmark to beatShort histories, new products
HZN-1best combination wins
Machine learningGradient boosting
Deep learningProprietary neural networks
Foundation modelsShort histories, new products
Causal modelsPrices, events and launches
Statistical modelsThe initial benchmark to beat

What you can do

A baseline you can trust, and explain.

Model selection

The best model for every demand pattern

  • HZN-1 runs models from five families against every product, customer and location pattern.
  • The best combination is picked automatically on accuracy, bias and stability, and stays current as new data arrives.
  • Why did the model change? Horizon shows which drivers mattered most and flags the forecasts it’s least confident in.
  • Planners adjust the forecast or test a what-if against the baseline, each with a reason code that guards against bias.
Model explainer
SKU-4471 · Retailer North · this week
Machine learningDeep learningCausalFoundationStatisticalMachine learningDeep learningCausalFoundationStatistical Tested against every product, customer and location pattern SELECTION SCORE AccuracyBiasStability Selected blend · updated this week WHAT DROVE THIS FORECAST PriceEventsSeasonalityCustomer38%27%21%14% FORECAST · NEXT 3 WEEKS W221,180W231,2401,320W241,210 Low confidence Reason code: customer intel

Demand behaviour

Demand behaviour, segments and hidden drivers explained

  • Products and customers clustered by volatility, growth, seasonality and value.
  • When demand moves, Deep Sensing pinpoints the factor behind it: price, an event, a customer, the market or supply.
  • Real demand change separated from noise, a data issue or a one-off event.
Demand segments
Products × customers · last 52 weeks
StableSeasonalGrowingVolatile Seasonal segment · weekly demand Real change ✓ NoiseData issue What moved it · W30 Event 55%Price 30%Cust. 15%

Recommendations

The next action, recommended before you ask

  • Each recommendation ranked by impact and priority, so the biggest risks surface first.
  • The next action suggested from how a segment is behaving right now.
  • Ask Ori, the planner’s assistant, a plain-language question about demand movement and get an explainable answer.
  • Outcomes fed back in, so recommendations improve every cycle.
Recommendations
Ranked by impact · this week
Seasonal-high trending above baselineHigh impactRetailer North launch moved to W12MediumLaunch SKU-5520 has a short historyMediumOnline channel: small positive biasLow NEXT ACTIONRaise baseline +8% · Seasonal-high, W22–W26AcceptAccepted ✓ Why is online demand down this week? It slowed after W19; the trend is stable. Outcome tracked · feeds next cycle ↻

Demand hierarchy

Aggregate and disaggregate demand across every hierarchy level

  • Demand aggregated from SKU up through category, customer and location, and disaggregated back down.
  • Dependent demand calculated from finished-goods requirements.
  • The granular number connected to S&OP, budget and leadership review.
Demand hierarchy
Weekly units · W22
Roll upSplit down
S&OPBudgetLeadership review +10% TotalCategory ACategory BSKU-101SKU-102SKU-103SKU-104SKU-105 3,0001,8001,2008006004007005003,3001,9801,320880660440770550 Component X · ×21,760Component Y · ×1880Dependent demand,from finished goods

Version compare

Any two versions, where they diverge

  • Compare any two saved sources (planner input, HZN-1 output or a prior version) over a custom date range.
  • Only items whose change exceeds your threshold are shown.
  • The largest change is called out, with trend lines for both versions and a variance chart by period.
Version compare
Planner input vs HZN-1 output · W14–W26
Planner inputHZN-1 output
Largest change · SKU-2210 · W20Variance by week W14W20W26 Show changesabove 2%above 5% SKU-2210 · Retailer North+14%SKU-4471 · Retailer North+9%SKU-0815 · Retailer South−3%SKU-3302 · Online+2%

Forecast accuracy

Accuracy against actuals, every period and lag

  • APE, MAPE and TAE plus directional bias, by lag or forecast horizon.
  • HZN-1’s forecast compared with the planner’s manual number, to see which performed better.
  • Accuracy by SKU, product line, warehouse or customer, not one aggregate number.
Forecast accuracy
Against actuals · last 13 weeks
HZN-1Planner
MAPEAPETAEBias Bias +3% · over-forecast 10%20% 12%15%18%24%16%19%21%22% Lag 1Lag 2Lag 3Lag 4 SKUProduct lineWarehouseCustomer Warehouse North11%Warehouse South14%Online17%Warehouse East24%
FAQ

Demand forecasting, answered.

How does Horizon build the baseline forecast?

HZN-1, Horizon's proprietary AI engine, runs five model families that compete, and the best combination wins per item. Results are deterministic: the same data gives the same answer.

What is Deep Sensing, and how does it change the forecast?

Deep Sensing reads signals beyond sales history, structured and unstructured: orders, promotions, CRM notes, customer emails and Excels, sales chats, call notes and external data. It links each signal to the items, customers and locations it affects and quantifies its impact.

How much more accurate is the forecast?

Horizon delivers 18–45% lower forecast error than traditional systems. Every forecast and every human input is scored against what really happened, so you can see what adds value and what doesn't.

At what level does Horizon forecast?

Across the hierarchy, from total and product group down to item × location and item × customer, so planning, sales and finance look at the same numbers at the level they work.

Do planners stay in control?

Yes. Language models read the text, the engines calculate the numbers and planners make the calls. Ori, the planner's assistant, ranks the decisions that matter by impact and explains each one, so only the key calls reach your planners.

How fast can we start?

With Horizon Instant, demand planning on your own data is up and running in less than 24 hours.

See your baseline on your own data.

Horizon Instant: up and running in less than 24 hours.