Walk into any supermarket at 8am and you will find a store manager deciding, in a few minutes and largely from memory, how much of each product to order for the week ahead. Multiply that moment across hundreds of SKUs, dozens of outlets and one distribution centre, and you have the reality of demand forecasting for retail in most of Southeast Asia today: thousands of small judgement calls, made quickly, with no way to check whether they were right until the shelf is either empty or overflowing. The technology to do better has existed for years. What has changed recently is that it now fits the budget, the data, and the operating conditions of ordinary retailers, not just global chains.
Three Levels of Inventory Maturity
Most retailers sit at one of three levels. The distance between them matters more than any software feature.
Level 1: The store manager’s instinct
Ordering is based on feel. An experienced manager knows the neighbourhood, remembers last Ramadan, and senses when something is moving faster than usual. This works surprisingly well for a single store, and it collapses completely at scale. The knowledge lives in one person’s head, it leaves when they leave, and it cannot be audited, compared or improved.
Level 2: Average and moving-average sales
More advanced retailers replace instinct with arithmetic: average sales over the last few weeks, sometimes weighted toward recent periods. This is a genuine step forward. It is also where most retailers stop, and it is where the damage quietly accumulates, because an average is a description of the past, not a prediction of the future.
Level 3: Predictive models
At this level, a model estimates demand per SKU and per location, then translates that estimate into a replenishment quantity. The difference is not that the model is smarter than the planner. It is that the model can hold hundreds of variables in view at once, apply the same logic to every SKU in every store, and do it again tomorrow without getting tired.
That gap between level two and level three is exactly where digitalplace.ai works as a technology partner, building AI-driven inventory optimization that fits how a retail network already runs.
Why Averages Fail
An average assumes that the next period will look like the last few. Retail rarely cooperates. Several patterns break the assumption completely, and each one is common enough to appear in any store, any week.
- Promotions. A discount does not lift demand by a stable percentage. The lift depends on depth, placement, competing offers and the category.
- Bulk orders. One large one-off purchase inflates the average for weeks afterwards, and the system keeps ordering for a customer who is not coming back.
- Seasonality. Festive cycles, payday effects and school calendars shift demand in ways that differ by location, not just by chain.
- Halo effects. A promotion on one item lifts related products that nobody flagged as promotional.
- Stockpiling. Customers buy ahead during an offer, then demand collapses in the following period — and the average reads that collapse as a genuine decline.
- Cannibalisation. Volume shifts to a substitute product on the same shelf, so the SKU looks like it is dying when the category is flat.
None of these is exotic. They happen constantly. The problem is that a moving average treats all of them as noise, and a human planner can only correct for two or three of them at a time across a handful of priority SKUs.
Research increasingly shows that machine learning and reinforcement learning can improve inventory decisions beyond traditional rule-based approaches such as (s,S) and Min–Max, particularly when demand is volatile and forecasting is connected directly to replenishment decisions. One comparative study evaluated reinforcement learning for dynamic inventory control, deep learning for multi-step demand forecasting, machine learning methods, and heuristic inventory policies using measures including service level, inventory cost and stockout rate. The results showed that more adaptive and hybrid approaches can deliver stronger inventory performance under changing demand conditions. (ScienceDirect)
A separate 2025 study took this further by combining transformer-based demand forecasting with multi-agent deep reinforcement learning to coordinate inventory decisions across a retail distribution network. In its experiments, the proposed framework achieved 18.2% lower forecast error and 23.5% fewer stockouts compared with the study’s state-of-the-art baselines, with particularly notable improvements during promotional periods and seasonal demand changes. (DOI)
Four Symptoms You Can See on the Floor
You do not need a data audit to know whether your inventory decisions are running on guesswork. Four symptoms show up in the aisles and in the P&L, and they tend to appear together.
Forecasting blind spots. Promotion, bulk-order and seasonality effects are very hard to predict by traditional formulas. When these events arrive, the plan is simply wrong, and everyone knows it before the data confirms it.
Stockouts on volatile and fast-moving SKUs. The items that matter most are the ones that run out first. Each empty facing is a lost sale and a small erosion of trust — the customer learns that your store is not reliable for that product.
Overstock on slow movers. Excess stock ties up working capital and drives holding and waste cost. It is less visible than a stockout because nothing appears to be wrong, which is precisely why it persists for months.
Traditional replenishment processes. Ordering based on a store manager’s feeling or on average sales cannot accurately predict demand, yet it still governs most reorder decisions in the market.
The combined result is the uncomfortable position many retailers find themselves in: working capital sits in the wrong stock while real demand goes unserved. Both halves of that sentence cost money, and they cost it at the same time.
What Actually Replaces Guesswork
Replacing guesswork is not about buying a forecast. It is about closing a loop, and the loop has four parts.
Data in. Outlets and distribution centres submit sales and stock data on a routine schedule. Routine matters more than real-time here — a scheduled upload works on low-spec devices and through unstable connections, which is the reality in most store networks.
Analysis. The model estimates what happens next, reading the demand signals behind each sale rather than the sales figure alone.
Daily result. Forecasts and replenishment quantities are produced per SKU and per location.
Execution. The distribution centre can use the result directly for store replenishment and for purchase orders to suppliers, with or without a human approval gate depending on how much autonomy the retailer is comfortable granting.
Just as importantly, the loop should cover the whole chain rather than one link in it. Many tools optimise the DC-to-outlet leg while supplier ordering stays manual, and the gains leak away upstream. Connected planning is the point: demand planning, inventory management and procurement stop operating as separate exercises. This is why the team at digitalplace.ai builds forecasting and replenishment to operate at every level of the supply chain — from supplier purchase quantities, to how stock is positioned across the distribution centre, to store-level replenishment matched to local demand.
Start With a Simulation, Not a Rollout
The honest way to evaluate any of this is to test it against your own history before committing to a deployment. Two to three years of transaction history is usually enough to capture seasonality, festive cycles and year-over-year trends, and a simulation on that data will show the expected movement in out-of-stock rates, days sales of inventory, sales and net margin.
The numbers from field deployments give a sense of what is realistic. In a controlled comparison across ten outlets served by a single distribution centre — current system versus model-driven replenishment — the results were roughly 25% of lost sales prevented, a 45% reduction in days sales of inventory, full automation of the ordering cycle, and a return on investment inside about four months. digitalplace.ai runs these simulations and deployments inside the retailer’s own environment, on-premise, so transaction and stock data never leaves the business, while model accuracy, out-of-stock rates, DSI and system uptime are monitored continuously to keep results on track. Results vary with implementation scope, but the mechanism behind them does not: better signals produce better orders, and better orders show up as freed working capital and fuller shelves.
The Bottom Line
Guesswork is not a failure of skill. It is what remains when the number of decisions exceeds the capacity of the people making them, which is the situation in every retail network above a certain size. The replacement is not a dashboard or a better spreadsheet. It is a model that reads demand signals per SKU and per location, an ordering cycle that runs without manual effort, and a feedback mechanism that tells you honestly whether the system is working.
Start by asking one question of your own operation: how many of this week’s orders could be explained by evidence rather than by experience? The answer usually makes the case on its own.
