Predictive analytics sounds like something for banks and airlines. For a distributor in Nagpur or a cold storage in Gondia it usually means three questions: what will sell next month, how much stock should I hold, and will there be enough cash on the 7th to pay salaries. This article covers what data you need, which simple models to start with, what a good forecast looks like and the mistakes we see.
What predictive analytics means for a mid-size business
Strip away the vocabulary and it is this: use what happened before to make a numbered guess about what happens next, then act on that guess. The guess will be wrong by some amount. The point is to be less wrong than "same as last month" or the owner's gut, and to know by how much.
Three areas pay back fastest for Indian MSMEs:
- Demand forecasting. How many units of each SKU will sell in each location over the next two to eight weeks.
- Stock optimisation. Given that demand and the supplier lead time, what reorder point and quantity keep you from running out without tying up cash.
- Cash-flow forecasting. Given outstanding invoices, each customer's payment habit and known outflows, what the bank balance looks like on each day of the next month.
Churn prediction, fraud detection and price optimisation are real but rarely the first thing to do; they need more data than most businesses of this size have.
The data you need, and where it usually is
Most businesses already have enough data. It is just scattered.
| Question | Data needed | Where it usually lives |
|---|---|---|
| Demand | Sales by SKU, by day, by location, for 18 to 24 months | Billing software, Tally exports, POS |
| Stock | Purchases, receipts, current stock, supplier lead times | Purchase register, stock ledger, supplier WhatsApp threads |
| Cash flow | Invoice dates, due dates, actual payment dates per customer | Accounts receivable, bank statements |
| Context | Festivals, holidays, price changes, promotions, weather | Calendars, memory, nowhere |
The last row matters more than people expect. A Diwali spike or a monsoon dip is obvious to a human and invisible to a model unless you record it. Twelve months of history is a minimum for anything seasonal; two years is comfortable.
Quality problems you will find: the same product under three names, stock adjustments used to hide losses, backdated invoices at month end, and customers who exist twice because of a spelling difference. Fixing these is most of the work.
Start with simple models
The temptation is to start with machine learning. Resist it. In our experience a business with clean data and a simple model beats a business with messy data and a sophisticated one, every time.
Moving averages and seasonal ratios
Average the last eight weeks of sales for each SKU, then multiply by a seasonal ratio (this week last year divided by the year's weekly average). This runs in a spreadsheet and is often within 15 to 25 percent of actual for steady products. It is a fine baseline and it is explainable to the store manager.
Exponential smoothing
Same idea, but recent weeks count more than older ones, and trend and seasonality are handled properly. Methods like Holt-Winters are built into most analytics libraries and are a sensible second step. They handle a product that is growing or fading better than a plain average.
Reorder point from forecast and lead time
Reorder point equals the forecast demand over the supplier lead time plus a safety stock. The safety stock depends on how variable demand is and how much you dislike stock-outs. This turns a forecast into a daily action: a list of what to order today.
Cash flow from customer behaviour
For each customer, compute the average days they actually take to pay, not the due date on the invoice. Apply that to open invoices and you have an expected inflow by day. Add known outflows and you have a cash curve. This is arithmetic, not AI, and it is the forecast most owners tell us they wish they had earlier.
A worked example
A distributor with 400 SKUs and two godowns pulls 18 months of invoice lines from billing. For the 80 items that make up most of the turnover, the weekly moving average with a seasonal ratio gives a four-week demand figure per godown. Supplier lead time is seven days, so the reorder point is one week of forecast demand plus a safety margin based on how much that item's weekly sales bounce around. Every Monday a list appears: item, godown, stock now, suggested order. The purchase manager edits it and sends the purchase orders. That is the whole system, and it can be built in a few weeks.
Only after these are running and trusted does it make sense to look at gradient-boosting models or neural networks, and mainly when you have hundreds of SKUs, many locations and external signals to include.
What "good" looks like
A forecast is not good because it is clever; it is good because someone changes a decision on the basis of it.
- A number for error. Track mean absolute percentage error per SKU group. Fast-moving, steady items might reach 10 to 15 percent; slow, lumpy items may stay at 40 percent or worse, and that is fine as long as you know it.
- Better than the baseline. Compare every model against "same as last period". If it does not win, do not deploy it.
- A person in the loop. The purchase manager sees the suggested order, adjusts it and the system records the adjustment. Over time you learn where the model is wrong and where the manager is, and both improve.
- A cadence. Weekly reforecast for stock, daily for cash. A forecast built once and never refreshed is a report, not a system.
- Visible in the tools people use. A suggested-order list in the billing system or a WhatsApp message at 8 am beats a dashboard nobody opens.
Pitfalls we see repeatedly
- Forecasting the total instead of the item. A company-wide revenue forecast is comforting and useless for ordering. Go to SKU and location.
- Ignoring stock-outs in the history. If you sold zero last July because you had none, the model learns that July is dead. Record lost sales, or at least flag stock-out days.
- Letting the model order automatically on day one. Run in suggestion mode for at least one season before anything is automatic.
- New products with no history. Borrow the curve of a similar product and correct as sales arrive. Do not pretend the model knows.
- Trusting one clean month. Indian demand is seasonal and festival-driven; validate across a full year before trusting it.
- Buying a platform before fixing the data. The tool is rarely the constraint.
- Forecasting collections from due dates. Invoices in India are rarely paid on the due date. A cash forecast built on due dates will be optimistic every month; build it on each customer's real payment history instead.
Rough costs and timelines
As a guide, a first predictive pipeline for a mid-size Indian business, covering data clean-up, a demand and reorder model for the main SKUs, and a dashboard or WhatsApp digest, typically lands between ₹1.5 and ₹6 lakh and takes six to twelve weeks. Adding cash-flow forecasting to existing accounts data is often the cheaper part. The real figure depends on how many data sources need connecting and how clean they are.
Our post on AI automation for small businesses covers what else can ride on the same data, and AI in ERP transformation shows where forecasting sits inside a wider system. If the stock side is your main pain, start with inventory management systems, because the forecast is only as good as the stock ledger beneath it.
Where SailRight fits
We build forecasting and analytics on the data a business already collects, usually starting from billing and stock exports, and the InsightRight module inside RightSutra covers the reporting side. Our AI and automation service starts with a data check before any model is promised. If you want to know whether your data is ready, get in touch.
- Predictive analytics
- Demand forecasting
- Inventory
- Cash flow
- Data
- MSME