"How many units should we order next month?" — For this question, many companies in Laos answer with "about the same as last year" or "a little more than usual, based on gut feeling." When stock runs short, they pay premium costs for emergency imports; when it runs over, capital sits idle in the warehouse. Demand forecasting AI tends to be seen as something that requires massive amounts of data and a dedicated specialist team — but that's not actually the case. With just three months of sales history and a hybrid approach combining statistical methods and AI, it's possible to start practical demand forecasting even without a data scientist on hand. This article walks through the concrete steps for enterprise companies in Laos to implement demand forecasting with minimal data.

Laos's supply chain presents unique challenges compared to other ASEAN countries. Beyond the geographical constraints of being a landlocked nation, the lack of a solid data infrastructure has long entrenched a reliance on intuition-based procurement. Here, we outline the structural challenges involved and the costs of leaving them unaddressed.

First, long lead times due to import dependency. Laos imports many goods—from consumer products to industrial materials—from Thailand, Vietnam, and China. Overland transport is subject to weather and infrastructure conditions, and it is not uncommon for delivery times to more than double during the rainy season. In an environment with long lead times, even a few days' miscalculation in ordering timing can result in stockouts.

Second, large swings in seasonal variation. Not only do consumption patterns shift dramatically between the rainy and dry seasons, but certain industries also see sharp spikes in demand during festive seasons such as the Lao New Year (Pi Mai Lao) and the That Luang Festival. Even experienced procurement managers find it difficult to anticipate these fluctuations based on past experience alone.

Third, fragmented data. Sales data in Excel, inventory in paper ledgers, and orders in a separate system—or existing only in the mind of the person in charge. Because data is not consolidated in one place, many companies cannot immediately answer even a basic question like "how many units of what sold last month."

Ordering without data inevitably causes two problems.

Excess inventory puts pressure on cash flow. Stock sitting in a warehouse is capital that cannot be moved. In high-interest-rate environments like Laos, the cost of holding inventory can reach 20–30% of the cost of goods sold. For food and consumables in particular, losses from disposal due to expired products cannot be ignored.

Stockouts create invisible opportunity losses. If products are not on the shelf, customers will turn to competitors. In B2B contexts, the situation is even more serious—delivery delays damage the trust of business partners and carry the risk of losing contracts.

The countermeasure most companies take is to "hold more safety stock," but this is nothing more than accepting the costs of excess inventory. Demand forecasting is a means of resolving this tradeoff between "gut instinct vs. cost."

When people hear "demand forecasting," many imagine feeding years of big data into a machine learning model. However, given the realities of Laos, that approach is overkill. Here, we introduce a "hybrid approach combining statistical methods and AI" that can be started with a small amount of data.

Weighted Moving Average (WMA) is a method that calculates an average by assigning "weights" to past sales data. By giving greater weight to more recent data, it produces forecasts that reflect current demand trends.

Specifically, weights are applied to the past three months of sales figures at a ratio of 0.5 : 0.3 : 0.2. Last month's data carries the most influence, while data from three months ago plays a supplementary role.

For example, if a product's sales over the past three months were 100 units, 80 units, and 120 units, the forecast for next month is calculated as follows:

120 × 0.5 + 80 × 0.3 + 100 × 0.2 = 104 units

Why WMA instead of advanced models like deep learning or ARIMA? There are three reasons.

• Minimal data requirements. Three months of data is enough to run it. Deep learning often requires several years of data.
• Results are easy to interpret. You can explain "why this number was reached" to management. It is not a black box.
• Virtually zero operational cost. No GPU or dedicated server is needed. It is simple enough to calculate in Excel.

WMA is strong for stable demand but cannot handle sudden fluctuations. New product launches, competitor price changes, government regulatory shifts——WMA cannot capture these "changes that haven't yet appeared in the numbers."

This is where LLMs (Large Language Models) play a supplementary role. When the WMA forecast results and sales history patterns are passed to an LLM, it returns analyses such as the following:

• Seasonality detection: "This product tends to see a 30% increase in demand every Q4."
• Trend identification: "Sales have grown 15% month-over-month for the past three months, indicating an upward trend."
• Anomaly warnings: "Last month's sales were 3x the historical average. This is likely a temporary spike."

The key design principle is that statistics are primary and LLM is supplementary. Even when LLM analysis is unavailable, the WMA forecast results can still be used as-is. Rather than over-relying on AI, the system delivers better predictions when AI is available——this "graceful degradation" design is especially important in environments where infrastructure is unstable.

Companies managing thousands of SKUs in inventory don't need to apply the same level of effort to forecasting every product. ABC/XYZ segmentation is a method that classifies products along two axes: "importance" and "demand stability."

ABC Classification (Importance by Sales Value)

XYZ Classification (Demand Stability)

By applying this classification, companies can adopt a strategy of leaving AX products (high-value and stable) to automated WMA forecasting, while concentrating LLM analysis and human review on AZ products (high-value and unstable). This allows limited resources to be focused on the products that truly require attention.

"I understand the logic, but how do I actually get started?"——From here, we will break down the actual implementation into 4 steps and explain each one.

The first step is to review the data you have on hand. You only need the following four types of data:

• Monthly sales volume (by product, covering the past 3 months or more)
• Current inventory quantity
• Lead time from suppliers (number of days from order placement to receipt)
• Order units (case units, pallet units, etc.)

You might be thinking, "Our data is in Excel, so it won't work," but there's no need to worry. As long as you can convert it to CSV, that's sufficient. It is far more valuable to start small with the imperfect data you have now than to spend six months striving for perfect data.

There is, however, one important caveat. Sales data recorded as zero during periods of stockouts does not mean "demand was zero." Products simply didn't sell because they weren't on the shelf. If you don't correct for this "censored demand," forecasts for out-of-stock items will come in lower than actual demand, leading to a vicious cycle of continued stockouts (the stockout loop). A practical correction method is to fill in the gaps using the average of sales data from before and after the stockout period.

All products do not need to be targeted for forecasting right away. Start by conducting an ABC analysis and picking up 10 to 20 A-rank products that account for the top 80% of sales revenue. Simply improving forecast accuracy for this product group will have a significant impact on overall inventory costs.

Run WMA demand forecasting for the selected products to project 3 months ahead. At this stage, do not aim for too much accuracy — there is inherent value in simply "producing a forecast." Each product is assigned a confidence score from 0 to 100, though scores in the 50–70 range will likely be common in the first run. It is normal behavior for products with greater demand variability to receive lower scores, so a low score does not mean the forecast is unusable.

Once the initial forecast is generated, conduct a backtest the following month by comparing it against actual results. By running the forecast under the assumption that data from a specific past month has "not yet been seen," then matching it against actual sales figures, you can quantitatively evaluate forecast accuracy.

Use MAPE (Mean Absolute Percentage Error) and WAPE (Weighted Absolute Percentage Error) as accuracy metrics.

• MAPE: The average error rate across individual products. Reveals the accuracy for each product.
• WAPE: Error rate weighted by sales value. Allows greater emphasis on accuracy for products with higher business impact.

In the Laos context, MAPE 20–30% is a realistic target. Since even large retail chains in developed countries typically achieve 15–20%, reaching below 30% in a data-limited environment can be considered sufficiently practical.

If accuracy is unsatisfactory, adjust the weighting parameters of the WMA. For example, for products with large seasonal fluctuations, you can automate optimization via grid search (exhaustive trial of all parameter combinations), such as increasing the weight of the most recent month further with values like [0.6, 0.25, 0.15].

Once forecast accuracy has stabilized, the forecast results are converted into a "recommended order quantity." The formula is straightforward:

Recommended Order Quantity = Inventory Cover Months × Monthly Forecast Demand − Current Inventory − Incoming Stock

For example, if the inventory cover is set to 2 months, monthly forecast demand is 100 units, current inventory is 80 units, and incoming stock is 0:

2 × 100 − 80 − 0 = 120 units

This quantity is then rounded to the nearest order unit (e.g., case = 24 units) to arrive at the final recommended order quantity.

For companies already using an ERP system (such as Odoo), the ideal approach is to automatically retrieve sales and inventory data via API and connect it to the forecasting pipeline. Even without an ERP, operations can be managed through monthly CSV uploads.

A weekly review process is essential for successful adoption. The procurement team reviews the gap between forecasts and actuals every week, and discusses the root causes of any anomalies. This repeated cycle drives the shift from a culture of relying on intuition to one of making decisions based on data.

The mechanism of demand forecasting itself is not difficult, but there are pitfalls that are easy to fall into during the implementation process.

The most common failure is setting unrealistic accuracy targets. There are times when people say, "If we're implementing AI, we'd like it to achieve 95% accuracy," but in demand forecasting, 95% accuracy is virtually unachievable except for certain everyday commodities with completely stable demand.

What matters is not absolute accuracy, but the improvement compared to "no forecast." "Stockouts decreased by 40% compared to when we were placing orders based on intuition," "Excess inventory decreased by 25%"——these kinds of improvements in business metrics are the right measure of the value of demand forecasting.

"I want to cover all SKUs if we're going to implement this" is an understandable sentiment, but it's counterproductive in the early stages. It is faster in the long run to first build a success story with the top 20 A-rank items, earn internal trust, and then gradually expand to B-rank and C-rank items — this approach ultimately accelerates company-wide rollout.

For C-rank products (the bottom 5% of sales), the return on investment of precise forecasting is simply low to begin with. They can be managed perfectly well with a fixed-quantity ordering system (placing a fixed order quantity whenever inventory falls below a certain level).

"First, let's build a data warehouse, complete the data cleansing, and then..." — if you plan this way, a year will pass before forecasting even begins. If you have three months of sales data in Excel, that's enough. Improving data quality on the fly will yield results far more quickly.

The WMA-based demand forecasting does not require statistical expertise. It is sufficient to be able to organize monthly sales data in Excel and understand the meaning of the formulas. The LLM enrichment component also requires parameter tuning but no model building, and can be used via API.

WMA can operate with 3 months of monthly sales data. However, 12 months is recommended to capture seasonal fluctuations, and 6 months is ideal for stabilizing accuracy. The system is designed to start with 3 months of data and improve in accuracy as more data accumulates.

Major ERPs including Odoo can retrieve sales reports and inventory data via API. It is possible to build an integration where the forecasting pipeline automatically pulls monthly sales performance and current inventory from the ERP, then feeds recommended order quantities back into the ERP's procurement module. For those not using an ERP, operations can be managed through CSV file import and export.

LLM enrichment plays a supplementary role, so the cost is extremely low. Even when analyzing monthly forecasts for 100 products with an LLM, the API usage fees amount to only a few dollars. The WMA calculation itself can be run on a standard server, with no GPU required. A phased approach is also possible: start by running WMA alone in the initial stage, then add LLM once improved accuracy becomes necessary.

The lack of data, IT talent shortages, and underdeveloped infrastructure that Lao businesses face are not "reasons why advanced AI cannot be used" — they are "reasons why a simple, robust approach is optimal."

With a hybrid approach combining weighted moving averages and LLM enrichment, you can start today with just three months of sales data and Excel. There is no need to cover every product. Begin with demand forecasting for the top 20 items by sales, improve accuracy through weekly reviews, and gradually spread that success throughout the organization.

Supply chain optimization is not something completed through a single large-scale investment. Starting with small forecasts and cultivating a culture of data-driven decision-making is the first step toward fundamentally transforming the competitiveness of Lao businesses.

Our company provides demand forecasting solutions to support supply chain optimization for Lao businesses. We offer end-to-end support, from integration with existing ERP systems to post-implementation operational adoption. We would love to first hear about the challenges you are currently facing.