Make Demand Forecasting Coherent: 4 Methods for Intermittent Cases

Demand forecasting estimates future customer demand so teams can set inventory, staffing, and production levels to hit service targets without carrying excess cost. A retailer uses it to decide how many units to hold before a seasonal peak, and a manufacturer uses it to schedule production runs weeks ahead. Practitioners judge forecast quality with metrics like MASE, and this guide, from BotiqueAI, walks through the methods that hold up in practice.

TL;DR:

• Intermittent demand, common in slow-moving products, requires specialized methods like mTSB, which outperform standard smoothing techniques on sparse data.
• Hierarchical demand data must be reconciled with methods such as optimal reconciliation or middle-out to ensure consistent forecasts across products, locations, and time periods.
• Forecast accuracy should be monitored with metrics like MASE, bias, and RMSE, and metrics must be used to inform safety stock and inventory decisions.
• Automating forecast integration into inventory and ordering systems is essential for operational use and continuous improvement and should follow GDPR protocols from the start.
• External market shifts and external shocks require ongoing bias monitoring and potentially causal models to incorporate real-time external information.

Table of Contents

• Choosing among the core forecasting method families
• Why intermittent demand breaks standard smoothing methods
• Making forecasts add up across products, locations, and time
• Building the data and feature foundation for production forecasting
• Turning forecast error into inventory and service decisions
• Fixing the mistakes that quietly wreck forecast accuracy
• How BotiqueAI turns forecasts into operational automation
• Where demand forecasting earns its keep across industries
• How external shocks and market shifts move your forecast
• Matching the model to the context, not the trend
• Planning for demand you can’t pin to a single number
• Connecting forecasts to inventory and production systems
• Handling privacy and fairness in demand data
• What’s actually changing in demand forecasting practice
• Build your forecasting pilot into a working automation
• Sources
• FAQ

Choosing among the core forecasting method families

Every forecasting approach trades off data requirements against accuracy and explainability, and the right pick depends on what you’re forecasting, not on which method is trendiest.

Qualitative or judgmental methods rely on expert opinion, sales team input, or structured techniques like the Delphi method. They fit new product launches or markets with no sales history, where no time series exists to model. A regional sales lead’s read on a competitor’s pricing move often beats a statistical model with zero historical signal to learn from.

Classical time-series models, chiefly exponential smoothing (ETS) and ARIMA, remain the workhorses for established products with steady sales histories. ETS handles trend and seasonality with relatively little data, often 2 to 3 years of monthly history, and produces forecasts that are easy to explain to a planning committee. ARIMA needs more data and more statistical care in setting parameters, but it can capture autocorrelation patterns that simple smoothing misses. Both assume the future looks broadly like the recent past, which is their main weakness during structural shifts.

Causal or regression models bring in exogenous drivers: price, promotions, weather, or macroeconomic indicators. They matter most when demand doesn’t move on its own internal rhythm but reacts to something you control or can observe, like a promotional calendar or a price change. A grocery chain forecasting cereal sales without a promo flag will misfire every time a discount runs, because the model has no way to explain the spike.

Machine learning models such as gradient boosting or neural networks add value when you have many related series, rich features, and enough history to avoid overfitting; a company forecasting thousands of SKUs across dozens of stores can let a single model learn shared patterns that a series-by-series ETS approach would miss. The risk is opacity: a boosted tree ensemble is hard to explain to a finance team asking why next quarter’s number moved, and these models can quietly learn spurious correlations from limited training windows.

Decision rules that hold up across industries:

• Short horizon, dense data, one product: ETS or ARIMA usually wins on speed and clarity.
• Long horizon or new product with no history: start qualitative, blend in statistical methods as data accumulates.
• Demand reacts to price or promotions: add a causal or regression layer regardless of the base method.
• Thousands of related series with rich shared features: machine learning earns its complexity.
• Stakeholders need to see why a number moved: favor ETS, ARIMA, or regression over opaque ML.

Pro Tip: Run a simple ETS baseline alongside any machine learning model. If the fancier model can’t beat it by a meaningful margin, the added complexity isn’t paying for itself.

Why intermittent demand breaks standard smoothing methods

Intermittent, or sparse, demand shows up constantly in spare parts, medical supplies, and slow-moving SKUs in retail, where a product might sell zero units for weeks and then move five at once. Standard smoothing methods like SES and ETS assume demand arrives in a roughly continuous stream, so they tend to produce forecasts that drift toward a flat average and either overstock during dead weeks or understock right before a spike.

The Teunter-Syntetos-Babai (TSB) method addresses this by separately estimating the probability that demand occurs in a given period and the size of demand when it does occur, rather than smoothing the raw series as one signal. A modified version, mTSB, refines that estimation further. In an empirical study of 1,718 medicine products, mTSB achieved a low mean error, outperforming several benchmark methods on the same dataset.

mTSB’s mean error on 1,718 medicine products was the strongest result among the methods tested in that study, suggesting that intermittent-specific methods meaningfully outperform generic smoothing on sparse series.

Practical notes for adopting intermittent-specific methods:

• Test for intermittency first: if more than a third of periods show zero demand, standard ETS is likely to underperform.
• Consider temporal aggregation (weekly or monthly buckets instead of daily) before jumping to a specialized method, since aggregation alone smooths out some sparsity.
• Tune TSB or mTSB parameters on a holdout period specific to the intermittent SKUs rather than reusing parameters fit on fast-moving products.
• Reserve intermittent-specific methods for the SKUs that actually need them: applying mTSB across an entire catalog adds complexity without benefit for fast movers.

Making forecasts add up across products, locations, and time

A demand plan is only operationally useful when the numbers are coherent: store-level forecasts have to sum to the regional forecast, and weekly forecasts have to sum to the monthly one. Structural hierarchies group by product, location, or customer; temporal hierarchies group the same series by day, week, month, or quarter. When these don’t reconcile, a warehouse manager and a finance planner end up working from different numbers for the same business.

Four approaches handle this coherence problem, each with a different trade-off:

• Bottom-up: forecast at the most granular level and sum upward, which preserves local detail but can accumulate noise across thousands of SKUs.
• Top-down: forecast the aggregate and disaggregate down, which is simple and stable but can miss local patterns like a single store’s unusual seasonality.
• Middle-out: forecast at an intermediate level, then reconcile both up and down, balancing the two failure modes above.
• Optimal reconciliation: use the full hierarchy simultaneously and adjust every level statistically so all series remain coherent while minimizing overall error.

Optimal reconciliation methods often outperform naive bottom-up or top-down approaches, though they require more statistical infrastructure, typically a dedicated reconciliation library and staff comfortable with matrix-based forecasting tools. Teams with a small hierarchy and limited technical bench often do better with a well-run middle-out process than with optimal reconciliation implemented poorly. The right choice depends less on theoretical accuracy gains and more on whether the team can maintain the approach month after month.

Building the data and feature foundation for production forecasting

Most forecasting failures trace back to data problems, not model choice. Before comparing algorithms, get the inputs right.

1. Set your minimum data bar: at least 2 years of history at the granularity you plan to forecast, whether that’s daily transactions or weekly aggregates, and a clear separation between gross sales and returns.
2. Build calendar and event features: holidays, promotional flags, and day-of-week indicators explain a large share of the variance that raw sales history alone can’t capture.
3. Add lag features and simple price or promotion proxies, such as a discount percentage or a binary promo flag, so causal effects are visible to the model instead of buried in noise.
4. Clean before modeling: treat extreme outliers (a one-time bulk order, a data entry error), decide explicitly how to handle true zero-demand periods versus missing records, and reconcile returns and adjustments so they don’t get counted as new demand.
5. Match your tooling to your stage: a spreadsheet pilot proves the concept, open-source libraries like statsmodels or Prophet handle a moderate SKU count, and a production pipeline with scheduled retraining and monitoring is what sustains accuracy once the forecast feeds real inventory decisions.

Pro Tip: Log every forecast alongside the actual outcome from day one, even during a spreadsheet pilot. That history becomes your backtest dataset later, and rebuilding it retroactively is far harder than capturing it as you go.

Turning forecast error into inventory and service decisions

A forecast is only as useful as your ability to measure how wrong it was and act on that information. Four metrics cover most practical needs. MASE (mean absolute scaled error) compares your model against a naive baseline and works well across series of different scales, which makes it a solid default for comparing methods. MAPE (mean absolute percentage error) is intuitive for stakeholders but breaks down on intermittent series with many zero-demand periods. RMSE penalizes large misses more heavily, useful when big errors are disproportionately costly. Bias, or mean error (ME), tells you whether you’re systematically over- or under-forecasting rather than just how far off you are on average.

Rolling-origin backtests, where you repeatedly retrain on an expanding window and test on the next period, give a far more honest accuracy read than a single train-test split, because they show how the model performs across different seasons and demand regimes. Stratify your holdout sets by product category or intermittency level so a strong average score doesn’t hide poor performance on a specific subset.

• Persistent positive bias signals you need more safety stock or a model correction, not just a tighter service-level target.
• High MASE relative to a naive benchmark means the model isn’t earning its complexity and a simpler method may serve better.
• Metrics should feed a documented safety-stock formula, not sit in a report nobody acts on.
• Assign a named owner for forecast accuracy review on a fixed cadence, monthly at minimum for fast-moving categories.

Industry guidance on demand planning stresses analyzing metrics by product, customer, and location rather than a single blended company-wide number, since a good aggregate score can mask poor performance in the segments that matter most for cost control.

Fixing the mistakes that quietly wreck forecast accuracy

Most forecast failures repeat a small set of patterns, and each has a fast diagnostic.

• Overfitting and data leakage: a model that looks excellent in testing but degrades quickly in production often had future information leak into training, such as a promo flag set after the fact. Test by checking whether accuracy holds on a strictly out-of-time holdout.
• Aggregation mismatches: forecasting in fiscal weeks while inventory systems run on calendar months creates silent misalignment. Audit that every system in the pipeline uses the same calendar definition.
• Promotion handling errors: treating a promo period as normal demand inflates the baseline for future forecasts. Flag and separate promotional uplift explicitly rather than smoothing it into the trend.
• Post-hoc bias: adjusting forecasts after seeing results, without documenting why, erodes trust in the numbers over time.

When accuracy drops, simplify the model, recheck the data pipeline for a broken feed, and reevaluate features before assuming the algorithm itself is the problem.

How BotiqueAI turns forecasts into operational automation

A forecast that sits in a spreadsheet doesn’t move inventory. Some agencies build the automation layer that connects a forecast output to the systems that act on it: a low-stock alert routed to a notification platform, a reorder trigger sent to a messaging workflow, or a dashboard that flags SKUs drifting outside their expected demand band. That integration work, connecting a statistical model to a CRM, ERP, or e-commerce backend, is where forecasting projects tend to stall without dedicated automation support.

Deployments follow GDPR-aware data handling practices, since demand data often touches customer transaction records that fall under standard data-protection obligations. Details on author background and specific client case studies are not publicly listed here, but the underlying approach mirrors the same production-readiness standard applied across BotiqueAI’s automation and integration work.

Where demand forecasting earns its keep across industries

In supply chain operations, forecasting drives raw material procurement and production scheduling, letting a manufacturer commit to supplier orders weeks before demand materializes instead of reacting to shortages after the fact. In retail, forecasts set store-level replenishment quantities and inform markdown timing, directly shaping how much working capital sits in unsold inventory. E-commerce operations lean on the same forecasts to plan warehouse staffing around predictable order spikes, and pairing that signal with product recommendation systems lets a storefront surface the right items before a demand wave hits.

In finance, demand forecasts feed revenue projections and cash flow planning, since a sales forecast is, in effect, a revenue forecast one layer removed. A finance team building next quarter’s budget without a demand-planning input is essentially guessing at the same number operations teams are actively modeling. Retail automation projects increasingly link the two: a recent retail automation case shows forecasts feeding directly into fulfillment and recommendation flows rather than sitting as a standalone report.

Each of these applications shares one requirement: the forecast has to reach the system that acts on it, whether that’s a purchase order, a staffing schedule, or a budget line.

How external shocks and market shifts move your forecast

No demand model operates in a vacuum. Price changes, competitor moves, macroeconomic swings, and shifting consumer sentiment all shift demand in ways that a purely historical model can’t anticipate on its own. A causal or regression layer helps here, but only when the external factor is actually captured as a feature rather than left implicit in the sales history.

Market trends compound this further. A category experiencing a structural shift, like a permanent move toward online grocery ordering, breaks the assumption that recent history predicts near-future behavior, which is the core assumption behind ETS and ARIMA. Models trained on pre-shift data will systematically miss the new baseline until enough new data accumulates to retrain on.

Practical response: monitor forecast bias by segment on a rolling basis, and treat a sudden, sustained bias shift as a signal to investigate an external cause before assuming the model itself has degraded. Circular supply-chain shifts add another layer of external pressure. As France Supply Chain and ADEME describe, circular and slow-logistics models introduce returned and repositioned goods as a second demand-relevant flow, one that traditional forward-demand models were never built to capture.

Matching the model to the context, not the trend

Model selection should start from three questions: how much history do you have, how volatile is the demand pattern, and who needs to understand the output. A fast-moving retail SKU with three years of clean weekly sales and no major structural breaks is a good fit for ETS or ARIMA. A newly launched product with a handful of comparable historical launches calls for a causal model built around those analogs, or judgmental input until enough data accumulates.

Machine learning earns its place only when you have breadth: hundreds or thousands of related series where a shared model can learn cross-series patterns that a one-at-a-time approach would miss. Applying it to a single, low-volume series usually just adds overfitting risk without a corresponding accuracy gain.

Explainability requirements often override raw accuracy in the selection decision. A finance team that needs to justify a demand number to auditors will tolerate a slightly less accurate ETS model over a marginally better but opaque neural network, because the cost of not being able to explain a number can exceed the cost of a small accuracy gap. Intermittency changes the calculus again: once a series shows substantial sparsity, TSB or mTSB-family methods generally outperform generic smoothing regardless of how much data you have, because the underlying assumption behind smoothing methods doesn’t hold for that kind of series.

Planning for demand you can’t pin to a single number

Point forecasts hide the range of outcomes that’s actually plausible, and scenario planning fills that gap by modeling a small set of distinct futures rather than one best guess. A typical setup runs a base case alongside an upside scenario (a successful promotion, faster-than-expected market growth) and a downside scenario (a supply disruption, a demand shock from a competitor’s move).

Probabilistic forecasting takes this further by producing a full distribution, or at least key quantiles, instead of discrete scenarios, which lets planners size safety stock against a specific service-level target rather than a single expected value. This matters most for products where the cost of being wrong is asymmetric: stocking out on a critical medical supply carries a different cost than overstocking a low-margin commodity item, and a single point forecast treats both the same way.

Building this into a planning process doesn’t require abandoning point forecasts. Most operations teams layer scenario ranges on top of an existing point-forecast model, using the point forecast as the base case and building upside and downside bands from historical volatility or explicit business assumptions about a known risk.

Connecting forecasts to inventory and production systems

A forecast that never reaches an ERP or inventory system stays a spreadsheet exercise. The practical integration point is usually the reorder or production-planning module: forecasted demand feeds directly into a reorder-point calculation or a material requirements plan, and the system generates purchase orders or production schedules from that number automatically rather than waiting for a planner to transcribe it.

This integration also closes the feedback loop that makes forecasts improve over time. When a forecast drives an actual order and that order’s outcome is measurable, the resulting accuracy data flows back into the model’s evaluation, rather than living in a separate reporting process disconnected from what actually happened.

The technical pattern that works in practice: a scheduled job pulls the latest forecast, applies a business rule (minimum order quantity, lead time buffer), and writes the result into the inventory or production system’s native format, whether that’s an API call to an ERP or a scheduled file drop. Building this connective layer is often the difference between a forecasting project that gets used and one that gets read once and shelved. BotiqueAI’s automation work with n8n and Make is built around exactly this kind of system-to-system handoff.

Handling privacy and fairness in demand data

Demand forecasting runs on transaction-level data that often includes customer purchase histories, which puts it squarely inside standard data-protection obligations wherever that data includes identifiable customer information. Aggregating to SKU or category level before modeling, rather than working from raw customer-linked records, reduces exposure while usually preserving all the signal a demand model actually needs.

Bias is a subtler risk. A forecast trained predominantly on data from one customer segment or region can systematically underserved another, for instance under-forecasting demand in a lower-volume market simply because it’s underrepresented in the training history, which then perpetuates the underinvestment. Reviewing forecast accuracy by segment, not just in aggregate, is the most direct way to catch this before it compounds into a service gap.

Governance practices worth setting from the start: document what data feeds each model, restrict access to identifiable records to the people who actually need them, and retain a clear audit trail of when and why a forecast was manually overridden. None of this requires slowing down a forecasting program, but skipping it tends to surface as a compliance problem later, at a much higher cost than building it in from day one.

What’s actually changing in demand forecasting practice

Circular supply chains are the trend worth watching most closely. Once a business plans for returned or repositioned goods alongside forward demand, it’s forecasting two linked but distinct flows, and most existing tooling was built for one. That gap calls for new metrics, not just a new model.

AI adds real value only when the data underneath it is complete, particularly supplier-tier data several links removed from the point of sale. Where that completeness is missing, added model complexity mostly amplifies noise rather than signal.

For teams deciding where to spend the next budget cycle: pilot a probabilistic forecast on one high-stakes category, add basic reconciliation between two hierarchy levels you already track, and audit supplier-tier data quality before adding any new model.

— Botiqueai

Build your forecasting pilot into a working automation

Most demand forecasting projects stall at the report stage: a model gets built, a chart gets shared in a meeting, and nothing downstream actually changes. BotiqueAI closes that gap by building the automation layer between a forecast and the systems that act on it, so a demand signal turns into a reorder trigger, a restocking alert, or a dashboard update without a planner manually re-entering numbers.

Work typically starts as a scoped pilot rather than a long commitment, with GDPR-aware data handling built in from the first integration. Whether the forecast needs to reach a Shopify inventory system, a WhatsApp Business alert flow, or an internal dashboard, the integration work follows the same pattern: connect the model’s output to the system that needs to act on it.

• Scope a pilot around one product line or category before expanding to the full catalog.
• Connect forecast outputs directly to existing inventory or CRM systems instead of a standalone report.
• Keep data handling GDPR-aware from the first integration, not as an afterthought.

Explore custom AI automation built around your existing forecasting and inventory workflow.

Sources

• A modified Teunter-Syntetos-Babai method for intermittent demand forecasting
• Hierarchical forecasting improvements and probabilistic reconciliation (arXiv)
• Demand and supply planning — ASCM topics page

FAQ

Qu’est-ce qu’une prévision ?

Une prévision est une estimation quantifiée d’un événement futur, comme le volume de ventes d’un produit, fondée sur des données historiques et des méthodes statistiques ou de jugement. Elle diffère d’une simple supposition par sa méthode explicite et sa capacité à être testée contre des résultats réels.

Qu’est-ce que l’analyse de la demande ?

L’analyse de la demande consiste à examiner les facteurs qui influencent les ventes d’un produit, comme le prix, la saisonnalité ou les promotions, afin de comprendre pourquoi la demande varie. Elle précède souvent la prévision, car elle identifie les variables à intégrer dans un modèle causal.

Quelle est la différence entre prévision et prédiction ?

Les deux termes sont proches, mais la prévision (forecasting) s’applique généralement à une série temporelle avec un horizon défini, comme les ventes du mois prochain. La prédiction (prediction) est un terme plus large qui couvre aussi des estimations ponctuelles, hors contexte temporel, comme la probabilité qu’un client donné achète un produit.

Quelles sont les méthodes de prévision ?

Les méthodes se répartissent en quatre grandes familles: qualitatives (jugement d’expert), séries temporelles classiques (ETS, ARIMA), causales ou de régression (intégrant prix et promotions), et apprentissage automatique pour les catalogues à forte densité de séries. Le choix dépend de l’horizon, du volume de données disponible et du besoin d’explicabilité.

Comment gérer une demande intermittente ou sporadique ?

Une demande intermittente, fréquente pour les pièces détachées ou les produits médicaux, se gère mieux avec des méthodes dédiées comme TSB ou sa version modifiée mTSB plutôt qu’avec un lissage exponentiel classique. Une étude empirique a montré que mTSB atteignait une erreur moyenne d’environ 0,07, surpassant six autres méthodes testées.

Recommended

• Cut Downtime 30–50% With Predictive Maintenance for Manufacturers