From statistics to decision
Many companies still forecast demand using simple averages and basic rules. Those numbers end up being used as the basis for important decisions on inventory, production and materials.
The problem is that this approach ignores relevant patterns. Seasonality, trend, events and behavior changes simply don't show up.
When a well-applied predictive model comes in, it starts capturing those signals and improves accuracy consistently.
And this isn't a technical detail. Accuracy has direct impact on results.
Each incremental gain in forecast accuracy reduces excess inventory, avoids stockouts and improves service level. For every 1% improvement in forecast accuracy, two widely recognized market references help us size the impact:
IBF (Institute of Business Forecasting)
0,20%
increase in profitability for every 1% improvement
McKinsey & Company
0,50%
reduction in total inventory cost for every 1% improvement
See details in: Forecast Accuracy and Financial Impact
But even that doesn't solve the whole problem. Statistical forecasting should be just the starting point. If it isn't reviewed, contextualized and connected to the business, it remains just a number.
It's collaboration that turns forecast into plan.
When sales, operations and finance take part in the process, the forecast stops being an estimate and starts guiding real decisions.
This article shows how to structure demand planning end-to-end: from data and models to building an integrated plan connected to the financial result.
TL;DR
- Demand planning isn't just forecasting: it's deciding, based on clean data, business context and consensus across teams.
- The process has 5 stages: data treatment, enrichment, analysis, models and collaboration.
- There's no single model that always works: test several approaches and pick the one that performs best in a hold-out window, series by series.
- Measure Forecast Value Added (FVA): compare accuracy before and after manual adjustments to see where collaboration actually adds value.
- Explainability turns forecast into defensible decision: trend, seasonality, events and external variables decomposed.
- Real value shows up when the forecast becomes a plan consolidated across sales, operations and finance.
How Demand Planning works in NPLAN
Demand planning isn't an isolated step. It's a structured process that starts with data preparation, goes through analysis and statistical models, and ends in collaborative decision-making. Below is the full view of the process.
Each step below is detailed with practical examples and real platform interactions.
From data inputs to adjusted forecast
Demand drivers feed the model, the network generates the forecast with an uncertainty band, and the expert applies fine-tuning.
Data Treatment
What happens before the forecast.
How to structure the data hierarchy in planning
Hierarchy is the first step in demand planning. Before talking about forecasting, you need to define how data will be organized.
This structure is what enables navigating between aggregated and detailed views, filtering information, collaborating across areas and making decisions at different levels. Without a well-defined hierarchy, planning becomes a collection of numbers without context.
1. Source data
| Company | Region | Sales Channel | Customer | Segment | Brand |
|---|---|---|---|---|---|
| Empresa 1 | Nordeste | Distribuidor | Cliente 1 | Farma | Marca A |
| Empresa 1 | Nordeste | Distribuidor | Cliente 2 | Farma | Marca B |
| Empresa 1 | Sul | Varejo | Cliente 3 | Construção | Marca A |
| Empresa 1 | Sul | Varejo | Cliente 4 | Construção | Marca C |
| Empresa 1 | Sudeste | E-commerce | Cliente 5 | Farma | Marca A |
| Empresa 1 | Sudeste | E-commerce | Cliente 6 | Cosméticos | Marca B |
| Empresa 1 | Sul | Distribuidor | Cliente 7 | Cosméticos | Marca C |
2. Configure hierarchy (drag the fields)
Available fields
Current hierarchy
3. Resulting navigation tree
This same tree becomes the side navigation of the tool.
This structure defines how planning will be navigated. From it the user can roll up to an aggregated view, drill down to detailed analysis, quickly filter any cut and organize collaboration across areas. This same hierarchy becomes the side navigation of the tool, used in every planning step.
Hierarchy isn’t just data organization. It defines how planning will be thought through.
Every analysis, model and decision that comes next depends on this structure. That’s why, before talking about forecasting, the first step is making sure data is organized correctly.
Nulls and Outliers: fixing the basics
Not every zero means absence of demand. Not every spike represents a real trend.
Historical data often carries issues like periods with no sales due to stockouts, integration errors and atypical sales. If these points are not treated, the model interprets everything as normal behavior.
How to handle it in practice
- Identify zeros that don't represent real absence of demand (out-of-stock items, integration failures). Tag those periods and replace them with an estimate based on similar weeks/months (interpolation or mean of comparable periods without stockouts).
- Detect outliers with simple statistical rules (values above 3 standard deviations from a moving average) or more robust methods like IQR. For every outlier ask: was it a real event (promo, holiday) or a data error?
- Keep an audit trail of adjustments: who changed it, when and why. Without a log, excessive manual edits start hurting forecast quality instead of improving it.
- Without this treatment, the model learns noise as if it were signal and produces unstable forecasts, with excess inventory or unnecessary shortages.
Simulator: nulls and outliers
Toggle treatments and see the effect on the historical series.
The goal is not to "clean" the data, but to make it representative of the real demand behavior.
Simulation: manual tags
Click a scenario to apply the tag and see the corresponding adjustment.
Tags don’t just change the value. They explain the why.
Comparison: forecast with and without treatment
Without treatment
With treatment
The model doesn’t fix bad data. It learns from it.
With a clean base, history finally reflects what actually happened. But history alone doesn't know about launches, campaigns or next quarter's events — that's where enrichment comes in.
Data Enrichment
Bringing business context into the plan.
Launches, substitutions and discontinuations
Not everything that affects demand is in the history. Portfolio changes, campaigns and future events must be incorporated explicitly. That’s what turns a statistical forecast into a realistic demand plan.
The product portfolio is not static. Items are discontinued, replaced or added to the mix. If this isn’t handled, the model keeps projecting demand for items that shouldn’t exist, or ignores new products with no history.
How to handle it in practice
- New products (NPI): use analogy. Transfer the demand pattern from similar existing items (same category, price band, channel) and adjust with the commercial expectation (expected volume in the first months).
- Discontinuations and substitutions: set a clear cutoff date. From that point, zero out the old item's projection and, if there's a successor, transfer part of the historical demand (for example, 70% migrates to the new SKU).
- Promotions and campaigns: separate the promo effect from the baseline. Compute the historical uplift (e.g., Black Friday added 35% above normal trend) and apply it only in future promo periods. Avoid treating promos as noise or as a recurring pattern — they are punctual events requiring conscious adjustment.
- Incorporate this context explicitly before running the model. That way the forecast looks forward instead of just extrapolating the past.
Product lifecycle
Pick a scenario and see how the plan reacts.
From the marked date on, the item no longer exists in the plan. The system stops projecting demand for it.
The model doesn’t guess portfolio changes. They must be informed.
Set up a campaign
Set the parameters and see the effect on the forecast or the history.
The campaign directly changes the future forecast in the selected period.
Campaigns aren’t noise. But they aren’t pattern either.
Without context, the model looks backward. With context, the plan looks forward.
With clean data and context in place, you can look at the series and ask what it's telling you: is there trend? seasonality? does any external variable explain the behavior?
Data Analysis
Understanding demand behavior.
Analysis: three questions before forecasting
After treating and enriching the data, the next step is to understand demand behavior. Before projecting the future, you need to answer three simple questions: is there trend? is there seasonality? is there any external variable explaining the behavior?
This analysis is what turns historical data into applicable intelligence.
How to handle it in practice
- Decompose the series into trend + seasonality + residual. Methods like STL or regression with monthly/weekly dummies help separate each component.
- Validate seasonality across multiple years. If the pattern shifts a lot year to year, it may be contaminated by non-recurring events (promos, stockouts) and needs to be cleaned.
- External variables (CPI, GDP, retail sales, weather, sector indices): test correlations with different lags and only include those that demonstrably improve accuracy in the validation window. Forcing weak-correlation variables usually hurts the forecast.
- Don't use every available variable. Use only those that add real explanatory power. More variables almost always means more noise.
Monthly Seasonality
Pick the years to identify recurring patterns (monthly average).
feb, jul, sep, oct
jan, mar, dec
This pattern is one of the main drivers of the forecast. Ignoring seasonality means assuming all months are equal — which is rarely true.
Trend Analysis
Real history with linear regression trend line.
Separating trend from seasonality is essential. Without it, the model can interpret growth as seasonality — or the other way around.
External Data Analysis
Compare demand against an external variable and adjust the lag.
Macroeconomic
Consumption / Retail
Economic activity
Not every external dataset helps. Forcing correlation where there is none is one of the most common mistakes.
Forecasting isn’t just projecting numbers. It’s understanding the behavior behind them.
Once you understand the behavior, picking a model stops being guesswork. Each pattern calls for a different approach — that's what defines the next chapter.
Forecasting Models
How the system picks the best forecast.
There’s no single model that always works
The debate about “which model to use” is one of the most common and one of the least relevant.
There is no model that always works.
Each series behaves differently. Some have clear trend, others are seasonal, others are intermittent.
Best practice is not to rely on a single model. Automatically test different approaches — exponential smoothing, ARIMA, models with external regressors, intermittent-demand methods — and pick the winner based on actual performance in a validation window (hold-out), series by series.
That completely changes the planner’s role: from technical executor to decision-maker.
Special case: intermittent demand
Some families — spare parts, MRO, low-rotation items — show intermittent demand: many weeks or months with zero sales, interspersed with unpredictable spikes. Traditional models (moving average, exponential smoothing) fail here because they assume continuous demand.
Specialized methods like Croston (which separates occurrence probability from order size), TSB or temporal aggregation approaches (ADIDA) usually perform better. Always test several methods and pick based on accuracy in a hold-out window.
How NPLAN picks the best model
Select a model to highlight its forecast and see test-period accuracy.
Computed as 100% − WMAPE over the 3 test months. The smaller the gap between forecast and actual, the higher the accuracy.
Train
Where models learn the historical pattern.
Test
Where models compete against actual data. This is where accuracy is measured.
Forecast
Where the winning model is applied.
The chosen model isn't the most sophisticated. It's the one that performs best on the test period — closest to actuals.
How to measure forecast quality
Each metric answers a different question.
Question
What’s the volume-weighted error?
What it measures
Percent error weighted by actual volume. Supply chain favorite.
How to read
Lower = better. Reflects financial impact.
There’s no single perfect metric. In practice, WMAPE is usually the most relevant for business decisions.
Comparing models in practice
Accuracy (%) by series and model. Star marks the winner.
| Series | MA3 | Auto ARIMA | Seasonal Naive | Croston | Prophet | Chronos | Times FM | ADIDA | Theta |
|---|---|---|---|---|---|---|---|---|---|
| Aggregated | 17% | 15% | 19% | 24% | 20% | 23% | 36% | 38% | 23% |
| 4001-01 | — | 17% | 45% | 23% | 43% | 31% | 47% | 26% | 31% |
| 4002-01 | 20% | 43% | 4% | 25% | 20% | 21% | 35% | 52% | 28% |
| 4003-01 | 11% | 15% | — | 6% | 3% | — | 7% | 37% | 2% |
| 4004-01 | — | — | — | — | — | 5% | 16% | 40% | 1% |
| 4005-01 | — | — | — | 3% | 13% | 7% | 16% | 20% | 6% |
| 4006-01 | — | 15% | 3% | 12% | 14% | — | 14% | 44% | 11% |
| 4007-01 | 32% | 24% | 40% | 34% | 42% | 45% | 56% | 52% | 38% |
| 4008-01 | 37% | 1% | 9% | 36% | — | 30% | 49% | 46% | 35% |
| 4009-01 | 8% | — | 39% | — | 22% | 17% | 36% | 20% | 11% |
Each series may have a different best model. This is expected and desirable.
Which models NPLAN uses
Grouped by type. More complex doesn’t mean better.
Simple models (baseline)
Useful as reference and good for stable series.
MA3
3-period moving average (configurable for N periods).
Seasonal Naive
Repeats same period from last year.
Statistical models
Capture trend and seasonality in a structured way.
Auto ARIMA
Automatic ARIMA selection.
Auto ETS
Automatic exponential smoothing.
Theta
Decomposes series into theta lines.
Dynamic Optimized Theta
Theta variant with dynamic trend capture.
Prophet
Facebook’s additive model with trend and seasonality.
Intermittent demand
For series with many zeros: spare parts, MRO.
Croston Classic
Specialized for intermittent demand.
ADIDA
Aggregation for sparse demand.
Machine Learning
Captures non-linear relationships. Needs data quality and volume.
XGB Regressor
XGBoost-based regression.
LightGBM
Fast, memory-efficient gradient boosting.
Foundation Models (AI)
Recent AI-based models. Complementary to classics.
Chronos
Amazon’s foundation model.
Times FM
Google’s foundation model.
NPLAN allows combining different models into an ensemble (weighted combination) and supports custom models tailored to each business reality.
Automatic best-model selection
NPLAN runs all active models, compares them and picks the best per series.
Runs all active models on the test window.
Compares performance using the chosen metric.
Applies the winner to the forecast. Manual override is allowed.
This removes specialist dependency and ensures process consistency. In many cases the final solution is an Ensemble (weighted combination of multiple models), and NPLAN is flexible enough to support even custom models tailored to each business reality.
It’s not the most sophisticated model that wins. It’s the model that best explains your business.
AI Explainability: understanding why the forecast is what it is
Generating a forecast isn’t enough. In supply chain, decisions need to be justified. When the forecast changes, the question isn’t just “how much,” but mainly “why.”
Without explanation, the forecast becomes a black box — hard to trust, hard to defend. A good practice is not only to compute the forecast but to show how it was built.
The system decomposes the forecast into its main components: trend (structural growth or decline), seasonality (recurring patterns), events (campaigns, stockouts, adjustments) and external data (market factors). The final forecast starts from a baseline and is adjusted by each of these factors.
Decomposing the forecast into factors
Adjust each factor and see how it contributes to the final forecast.
Automatic AI explanation
"The forecast started from a baseline of 300k units. Trend increased the forecast with moderate intensity, and seasonality increased the result in a strong way. Promotions and events contributed in a strong way (+30k). External factors reduced the forecast by 10k. Final result: 390k units."
Why this matters day-to-day
Explainability lets you justify decisions to leadership, align teams (sales, operations, finance), spot errors or inconsistencies and adjust scenarios with more confidence.
The system also flags inconsistent situations — high forecast without trend or seasonality, external impact without relevant correlation, excessive manual adjustments — preventing decisions based on misinterpretations.
Without explanation, a forecast is just a number with a mathematical look. With explanation, it becomes a defensible decision.
That’s what lets you discuss the plan with confidence, instead of just accepting or rejecting numbers.
An explained forecast still needs to become a plan. And a plan, in practice, depends on sales, operations and finance looking at the same number.
Collaboration: turning forecast into decision
Consensus across levels, real-time financial impact, KPIs and scenarios.
From forecast to decision
In most companies, the process breaks exactly here.
The forecast does get made. But it isn't consolidated across teams, doesn't connect with finance and doesn't reach the business as a clear decision.
This is where planning stops being technical and becomes collaborative. Sales, operations and finance work on the same base, with different views of the same plan.
Consolidation and split across levels
Adjust the family and watch the system split down to SKUs. Adjust the SKU and watch the family roll up — always consistent.
Raising volume can grow revenue — but it can also shrink margin, depending on cost and mix.
KPIs that support the decision
The main dashboards used in S&OP cycles. Pick one to see its goal.
Executive Summary
Quick context in minutes: growth, accuracy, drivers, risks and opportunities.
Scenarios: simulating before executing
Compare full plans side by side: optimistic, conservative, with promo and with supply constraint.
Each scenario generates a full plan: forecast, financials and KPIs. Lets you test decisions before taking risk.
A good demand plan isn't born ready. It's built in short cycles, where each area adjusts its view until the final number makes sense for the business as a whole.
How to measure the value of collaboration: Forecast Value Added (FVA)
- Compare the accuracy of the statistical baseline with the accuracy after the planner's adjustments, the commercial inputs and the S&OP consensus. The difference, positive or negative, is the FVA.
- In many processes, a significant share of manual adjustments hurts the result. FVA makes this visible and helps calibrate where human judgment really adds value.
- Use FVA at every step of the process (baseline → planner → sales → consensus) to pinpoint exactly where adjustments add or destroy value.
Best practices and metrics
How to measure accuracy, added value and real business impact.
Which metrics to use and when
It's not enough to measure accuracy with a single metric. Each indicator answers a different question, and using the wrong one can hide real problems — or create problems that don't exist.
| Metric | What it measures | When to use |
|---|---|---|
| MAPE | Mean absolute percentage error between forecast and actual. | Items with continuous demand and reasonable volume. Avoid when values near zero exist. |
| WMAPE | MAPE weighted by each SKU's volume. | Mixed portfolios with very different-sized SKUs. The default metric for aggregated views. |
| Bias | Signed mean difference (positive or negative) between forecast and actual. | Spot systematic bias: chronic under-forecast leads to stockouts, chronic over-forecast leads to excess inventory. |
| FVA | Accuracy gain or loss at each step of the adjustment process. | Evaluate whether human judgment and S&OP consensus are improving or hurting the forecast. |
Best practices that separate a good process from an excellent one
- Hierarchical reconciliation: forecasts at different levels (SKU, family, channel, region) must reconcile. The sum of SKUs has to match the family forecast. Top-down, bottom-up or optimal (MinT) reconciliation methods ensure this coherence.
- NPI tracking: new products require review every cycle. Compare the launch forecast with actuals, adjust the analogy used and refine the adoption curve.
- Forecast Value Added: measure the impact of each manual adjustment. If S&OP consensus on average hurts the forecast, rethink the process: the issue may be incentives (sales inflating to secure product) or lack of data.
- Real business impact: accuracy is a means, not an end. Also measure the effect on inventory (days of cover, working capital) and on service level (fill rate, OTIF). A better forecast that doesn't reduce inventory or improve service is, in practice, not being used.
- Clear cadence: set short, fixed cycles (weekly or monthly) with owners and deadlines. Without cadence, the process turns into a sporadic event and loses power as a decision tool.
What changes when the forecast becomes a decision
In the end, the point isn't hitting a number. It's understanding the impact of decisions before executing them — and adjusting the plan while there's still time.
Demand planning solves nothing if it stops at the forecast. The value shows up when that number becomes a decision aligned across teams and connected to the financial result.
It's this sequence — data, context, model, explanation and consensus — that structures a robust demand planning process.
See it running in practice
Everything you saw here is part of the daily routine of companies already using NPLAN to plan with more accuracy and speed.
More than concepts, the value is in seeing this process working with your own data.
Talk to our team and see how it works in practice.
Schedule a demoNext in the series: Inventory Policy
With demand forecast defined, the next step is to decide how much to keep in inventory to absorb demand and lead-time variability.
Read the next article