Building Demand Forecasting Models to Reduce Inventory

24/09/2026

Step 1: Diagnose Data and Identify Gaps

The Vietnamese retail market is witnessing a peculiar phenomenon: revenue is growing, yet net profit margins are shrinking. The cause is not raw material costs or marketing spend, but rather the excessive inventory we are holding. Approximately one-third of the F&B and retail businesses I have worked with admit they do not know exactly how many staff members they need next week or how much raw material will spoil in three days. This is the time to sit down and look directly at the data we have. Many executives believe that buying expensive AI software is the solution. That is a major mistake. If your input data is messy, AI will only produce more sophisticated errors.

The leadership perspective is critical here. You must clearly define that the ultimate goal is not to "have a pretty dashboard," but to reduce operational costs. Ask yourself: Where are we losing money? Is it electricity for overworked cold storage? Or wages for staff who are idle on weekdays but understaffed on holidays? The answers to these questions will determine how you design the system. Do not try to forecast everything at once. Start with a flagship product or a specific geographic region. For example, if you sell instant noodles, focus on a province with clear seasonal fluctuations before scaling nationwide.

Required Inputs

• Detailed sales data from the last 3-5 years (by day, by SKU, by point of sale).
• List of seasonal variables: holiday calendars, weather patterns, local events.
• Basic operational cost data (electricity, shift-based staff wages).

Actions to Take

Clean the data. Remove noise from anomalous events (such as pandemics or supply chain disruptions) if you do not want the model to learn the wrong patterns. Group products with similar behaviors. Do not treat each SKU as a separate problem; that is the fastest way to overwhelm your IT team.

Visible Output

A clean dataset with time variables labeled. You can clearly see the peaks and troughs in the annual sales cycle.

Signs the Step is Complete

The operations team confirms that the data accurately reflects reality on the factory floor or in the store. There are no negative numbers or large, unexplained data gaps.

Step 2: Build a Role-Based Forecasting Logic Framework

Once you have clean data, the next step is to redefine how we view forecasting. This is where the difference between a passive reporting system and a proactive resource management system lies. I usually break this process down into three main perspectives to ensure no one is left behind in the workflow.

Leadership Perspective: They do not care about algorithms. They care about cash flow and risk. The forecasting model must answer the question: "If I increase raw materials by 10%, how will net profit change?" Design "What-if" scenarios for them. For example, a scenario where unusually hot weather lasts for two weeks, or a scenario where a competitor significantly lowers prices. Leadership needs to see the impact on the balance sheet, not just the quantity of goods.

Operations Team Perspective: These are the people under the most pressure. They need a tool to help them schedule shifts and order raw materials without guessing. The model must output specific numbers: how many packages, how many kilograms of raw material, and how many staff members are needed for each shift. Most importantly, the system must provide early warnings when there is a mismatch between supply and demand. Do not let them discover a shortage only when the truck has already arrived at the warehouse. Flexibility is key, but flexibility based on data is sustainable.

IT Team Perspective: They are concerned about scalability and latency. With large datasets from multinational retail chains like those in Mexico or Thailand, real-time processing is a technical challenge. Ensure the model can run on a cloud platform to easily scale up as the market expands. Do not build a closed system that cannot integrate with existing ERP or WMS software.

Required Inputs

• Business scenarios from leadership.
• Current operational processes of the execution teams.
• Current IT infrastructure and cloud budget.

Actions to Take

Hold cross-functional discussions to agree on Key Performance Indicators (KPIs). Leadership might measure by ROI, operations by stockout rates, and IT by system response time. Find common ground. If departments measure differently, you will never reach consensus on results.

Visible Output

Clear Functional and Non-functional Specifications for the software development team.

Signs the Step is Complete

All three groups agree on the same set of KPIs and clearly understand their responsibilities in the forecasting lifecycle.

Step 3: Integrate Seasonal Variables and Exogenous Data

Many forecasting models rely solely on sales history. This is like driving while only looking in the rearview mirror. It will seem fine until another car overtakes you from the right. Seasonal variables and exogenous data are the factors you cannot see in the rearview mirror, yet they determine your direction.

Combining historical purchase data with seasonal variables is not simply adding a "Month" column to a spreadsheet. You need to understand causal relationships. For example, in the lubricant industry, where AIVISION has partnered with clients, we found that demand depends not only on rainy or sunny seasons but also on the vehicle maintenance schedules of logistics companies. Or in the instant noodle industry, a slight change in rice prices can affect consumer psychology and drive changes in demand for accompanying toppings.

Be careful with overfitting. If you put too many variables into the model, it will fit past data perfectly but fail miserably with future data. This is the most common trap. Start with variables that have clear economic significance. Weather, holidays, energy prices, PMI index. Then, add more complex variables like social media trends or actual weather data once the basic model is stable. Do not forget that in some markets like the Philippines, local cultural factors can be much stronger than general macro factors.

Required Inputs

• Historical and forecasted weather data.
• National and local event calendars.
• Macro data (inflation, exchange rates, energy prices).

Actions to Take

Analyze the correlation between exogenous variables and revenue. Remove variables with low correlation or those causing multicollinearity. Use methods like ARIMA with exogenous variables or Prophet if the data has clear periodicity. For businesses with complex supply chains, consider machine learning models like XGBoost or LightGBM to capture non-linear relationships.

Visible Output

Version 1.0 of the forecasting model with Accuracy evaluated on the test set.

Signs the Step is Complete

The Mean Absolute Percentage Error (MAPE) of the model on the test data is significantly lower than the current manual forecasting method. Typically, we aim for a reduction of about 15-20% compared to the baseline.

Step 4: Pilot Deployment and Measure Real-World Impact

This is the step where most AI projects fail. Not because the technology is poor, but because people are not ready to accept change. You cannot deploy a new resource management system without changing employees' working habits. Start with a small but representative pilot process. Choose a specific factory, warehouse, or retail region.

During this process, run two systems in parallel: the old system (manual forecasting) and the new system (AI forecasting). Compare results daily. Do not just look at the total numbers. Examine the exceptions. When the AI forecast is wrong, why? Is it due to incorrect input data, or does the model fail to handle a newly emerged variable? This is when the IT and operations teams must sit together to debug not just the code, but the business process.

A specific example: At a partner in the beer industry, we noticed that the initial model forecasted too low for weekends with sporting events. After adding the variable "major sporting event within a 50km radius," accuracy improved significantly. This demonstrates the value of continuously refining the model based on real-world feedback. Do not be afraid to admit that the model is wrong. A wrong model that is fixed quickly is far better than a model that is "correct" on paper but no one trusts.

Required Inputs

• Version 1.0 of the forecasting model.
• Operations team trained to use the new dashboard.
• Feedback loop from the field.

Actions to Take

Deploy an intuitive dashboard for the operations team. Clearly display forecasts, confidence intervals, and alerts. Hold short weekly meetings to review deviations. Record all feedback and add it to the development backlog.

Visible Output

A report comparing the performance of the old and new processes after a 3-month pilot deployment.

Signs the Step is Complete

The operations team proactively uses AI forecasts rather than viewing them as just a reference tool. Excess inventory in the pilot area decreases significantly, with a target reduction of about 20% compared to the baseline.

Step 5: Scale Up and Automate Processes

Once you have proven effectiveness in one area, the next step is to scale. But scaling does not mean copy-paste. Each region and market has its own characteristics. The Mexican market has a different supply chain structure than Thailand, and both differ from Vietnam. Prepare a scaling framework rather than a rigid script.

Automation is the keyword here. But be careful. Automating a wrong process will only make you wrong faster. Before automating, ensure the process is standardized and measurable. For example, instead of having staff manually input forecasts daily, set up automatic ordering rules based on safety stock levels calculated by the model. However, always keep a human-in-the-loop layer for major decisions. AI is good at processing large data, but humans are good at understanding context and diplomacy.

Think a bit further into the future. With the development of Agentic AI, future AI agents will not only forecast but also automatically perform actions such as negotiating with suppliers, adjusting production schedules, or optimizing delivery routes. This is the direction we are exploring with major partners. But that is a story for the future. Right now, focus on building a solid data foundation and an evidence-based culture within your organization.

Required Inputs

• Success results from the pilot phase.
• Budget for IT infrastructure expansion.
• Change management team to support training.

Actions to Take

Design a distributed system architecture to handle multi-region data. Automate repetitive tasks in purchasing and personnel processes. Establish a model monitoring mechanism to detect performance drift over time.

Visible Output

A forecasting system operating stably across the entire network, deeply integrated into ERP and WMS processes.

Signs the Step is Complete

Average operational costs per product unit decrease continuously. Staff can focus on creative and strategic tasks instead of struggling with manual forecasting tasks.

AIVISION helps enterprises turn AI into working systems. Explore our enterprise AI solutions, read more on the AIVISION blog, or talk to our team about your own use case.