How to perform demand forecasting with AI - A step by step guide

In this how-to guide and tutorial, you will use IBM Planning Analytics to generate demand forecasting with artificial intelligence (AI) driven insights.

Demand forecasting is the systematic process within supply chain management that uses historical data, market trends and external economic indicators for demand planning.

Predicting future customer demand is crucial for businesses to manage supply chain operations and inventory planning. With accurate data-driven forecasts, businesses can optimize inventory levels; stock levels are maintained to meet demand and ensure customer satisfaction (avoiding stockouts and shortages) without over-allocating capital to excess goods (minimizing overstocking).

AI demand forecasting takes the process further by implementing scalable AI-enabled forecasting tools and predictive analytics to estimate future demand. These systems can automate the analysis of thousands of variables simultaneously, identifying correlations that a human analyst might miss.

Robust demand predictions contribute to the wider organization’s budgeting and integrated financial planning. Informed with accurate demand forecasting, CFOs and finance teams can manage capital allocation in a cost-effective manner, ensuring cash flow is available for procurement when demand is expected to peak.

The “how” of forecasting is generally split into two categories: qualitative and quantitative methodologies.

Qualitative methods:

• Delphi method: A panel of experts provides independent expert opinions and forecasts. A facilitator summarizes their predictions and allows the experts to revise their views until a consensus is reached. This method is especially helpful for new products where no historical data exists.
• Market research: Uses customer surveys and focus groups to gauge customer needs and “intent to purchase.” While intent does not always result in action, it provides a pulse on market demand, customer behavior and sentiment that numbers cannot capture.

Quantitative methods:

A quantitative approach to forecasting is typically rooted in time series models, a type of machine learning model that analyzes chronological data to predict future values. Some time series methods include:

• Moving averages: This method calculates the average of a specific number of recent periods. By using the average, it smooths out random fluctuations but can lag behind rapid market changes.
• Exponential smoothing: This method is like the moving averages method but gives more weight to the most recent data points. It is generally more responsive to recent shifts in consumer behavior and market trends.
• Regression analysis: This approach analyzes the relationship between a dependent variable and one or more predictor or independent variables in temporal data. 

For our tutorial, we will use the statistical forecasting software included with Planning Analytics to create a baseline quantitative model.

For this walkthrough, you will need to set up an IBM account and register for a free Planning Analytics trial.

You cannot forecast effectively without knowing exactly what you are trying to achieve. Ambiguity at this stage of the demand forecasting process leads to “forecast drift,” where the data becomes too broad to be actionable.

First, identify the time frame that aligns with your decision-making cycle. If your raw material lead time is three months, for example, a weekly forecast might be too granular, while an annual forecast will be too imprecise.

• Short-term forecasting (typically < 1 year): This time frame is operational in nature, informing weekly production schedules, labor requirements and immediate inventory replenishment.
• Long-term forecasting (2–5+ years): This time frame informs business decisions regarding capital expenditure, such as building a new warehouse, investing in R&D or expanding into international territories.

Next, choose the right model for your specific need. Different types of demand forecasting models serve distinct strategic needs.

• Macro-level forecasting examines the broader market environment and shifting market conditions or volatility. It can consider external factors such as GDP growth, national unemployment rates, consumer price indices or other economic conditions.
• Micro-level forecasting drills down into specific business units, regions or customer segments. It focuses on how a specific brand — whether in retail or ecommerce — and its unique price points will perform within the larger macro environment.

Finally, to follow an effective demand forecasting and planning process, you must distinguish between two primary demand forecasting techniques.

• Passive demand forecasting relies solely on historical past sales data for future sales forecasting. It assumes that past patterns will continue without significant intervention or market shifts. This technique is best suited for stable businesses or products in the mature stage of their product lifecycle.
• Active demand forecasting accounts for aggressive growth plans, marketing campaigns and anticipated competitive responses. If you are managing new product launches or entering new markets, active forecasting is essential as it integrates your own strategic initiatives into the mathematical model. 

For our guide, we will implement a one-year, short-term, micro-level with hybrid (passive and active demand) forecasting scenarios. We will use 2025 historical data to predict 2026 sales data for three different products across global regions.

To access our example, on the Planning Analytics home page under “Your recommended tasks,” select “Update a demand plan using AI.”

The quality of your future demand output is entirely dependent on the quality of your input. This stage is often the most time-consuming but is nonnegotiable for generating accurate predictions.

First, collect your historical sales data. Extract data from your enterprise resource planning or point of sale systems to build comprehensive datasets.

For our walkthrough, you can see that the “Overview” page lists all the units we sold in 2025. It also lists the revenue, gross margin, operating costs and net income derived from those unit sales.

Next, if needed, data cleaning will help remove any noise from a dataset. For instance, if you had a one-time bulk order from a contract that won’t repeat, remove that data point. These outliers will otherwise skew your averages and lead to over-purchasing.

In our Planning Analytics demand forecasting example, we will proceed with the data as it stands without more cleaning.

Next, let’s identify some patterns in the data, including:

• Seasonality: Regular fluctuations that repeat across the temporal data
• Cycles: Longer-term fluctuations that don’t have a fixed period, often tied to the business cycle
• Trends: The general upward or downward direction of the data over time.

Staying on the “Overview” page, we can see that units sold peaked in March and July 2025. Also, there was a downward trend in the number of units sold in the fourth quarter. Per the notes, this decline was due to inventory shortages from higher-than-expected product demand causing sales disruptions in the second half of 2025.

Click “Next” to proceed.

Next, we can create our baseline forecast. This is our starting point—the mathematical prediction of what will happen if current trends continue.

Planning Analytics automatically generates the baseline statistical forecast for us. Let’s review and interpret the automated initial forecast and perform a “sanity check.” Reviewing the statistical forecast, we see peaks in March and July 2026, similar to the peaks from 2025.

Click “Next” to continue.

For our first adjustment, we receive important feedback from the sales team to address the unexpected increased consumer demand from 2025. They would like us to revise the forecast by adding 30,000 units to the US market for 2026. 

To do this revision on Planning Analytics, select the first demand forecast scenario, “DemandPlanScenario1,” from the Sandbox dropdown and “USA” from the Markets dropdown. Next, manually add ‘30000’ to the cell where the 2026 column and the “Sales/Marketing Adjustment” row meet. The 30,000 extra units will be distributed evenly across the 12 months to increase each monthly forecast by 2,500 units.

Let’s also review the AI-generated data analysis and insights we have in Planning Analytics for this first scenario.

Based on the AI-powered chart insights summary and key points, we get instant advanced analytics, including stats like standard deviation, average and median and other important key points from the forecast.

Click “Next” to move on to our second adjusted forecast.

For our second scenario, the sales and marketing team wants us to add 1,000 units to each month for 2026. We can automatically add 1,000 units to each month by entering ‘1000>’ in the cell for Jan 2026, “Sales/Marketing Adjustment.”

Also, the team is projecting a 20% growth month over month. We can include this adjustment by entering the command ‘grow20>’ in the same cell. Notice how each month increases 20% over the prior month. The new total “Sales/Marketing Adjustment” is 39,581 units for 2026.

Finally, the product management team anticipates even more sales, and they would like an extra adjustment of 2,000 total sales for 2026. We can input the ‘2000’ extra sales in the “Product Management Adjustment,” 2026 cell to allocate the 2,000 units evenly across all months. The new demand forecast for this scenario results in 60,019 total units for 2026, a more bullish outlook for the year compared to the prior scenario.

Feel free to review the AI-generated insights for the demand forecast visualization here too. Select “Next” when finished.

Once we’ve created the different demand forecasting plans, it’s time to share the scenarios with our stakeholders.

We’ve included feedback from both the sales and marketing, and product management teams in our projections. We should also make sure to communicate the different predictions with the finance, procurement and production teams. The procurement team can use these numbers to optimize its procurement plans and automate purchase orders, while the production team can integrate the forecasts to set labor shifts.

Continuing with our example, the next slide compares the three scenarios: the base model, “DemandPlanScenario1” and “DemandPlanScenario2.”

The “Sales, total costs and net income comparison” chart is especially helpful to get a snapshot of the forecasted profitability in the three different scenarios.

By clicking “Next” we can dive deeper in the final slide of our example, the financial comparison. Here, we can review the projected income and cash flow statements for 2026 for the three different scenarios.

As the year progresses, it is crucial that we measure our forecast accuracy. The standard metric to calculate accuracy is the mean absolute percentage error (MAPE). The goal is to reduce this percentage over time.

Formula: (Actual - Forecast) / Actual

We should also establish a feedback loop for continuous improvement; a forecast is a living document, not a static report. The sales and marketing team projected a 20% growth in unit sales for the more aggressive demand forecast scenario. As the year proceeds, how accurate was that scenario? We can use real-time data to update our forecast if necessary.

In this tutorial, you performed demand forecasting with Planning Analytics and AI.

You followed a structured approach of defining clear objectives, data collection, appropriate model selection, forecasting and monitoring to make well-informed decisions for your inventory management system.

To explore demand forecasting and other AI-enriched FP&A solutions, sign up for a free Planning Analytics trial.