Forecasting Demand Without a Data Science Team

Short answer: you can get a useful demand forecast from two or three years of your own sales history and a spreadsheet or a few lines of SQL. Start with last year's same period, adjust it with a moving baseline, and only reach for regression or machine learning when you can show the simple version is not good enough. Most of the value comes from doing the simple thing consistently and checking it against what actually happened.

This post walks through the methods I start with, how to tell whether a forecast is any good, where these methods stop working and when it is worth bringing in more help.

Start with the Decision the Forecast Feeds

A forecast is only useful if it changes a decision. Before you touch the data, write down which one:

• How much stock to order, and how far ahead
• How many staff to schedule each week
• How much cash you need before a slow season
• Which products to promote, and when

The decision sets the grain. Purchasing might need weekly units per product, eight weeks out. Staffing might need daily orders, two weeks out. Cash planning might need monthly revenue, six months out. A forecast at the wrong grain will not get used.

The Data You Already Have Is Usually Enough

For most retail, ecommerce and service businesses, the starting data is an order or transaction history with a date, a product or service, a quantity and a price. Two full years is the minimum to see a seasonal pattern repeat. Three is better.

Clean it before you forecast it. Remove test orders and internal transactions, decide how to treat refunds, and mark the weeks that were not normal: a stockout, a site outage, a one-time bulk order, a closure. If you skip this, the forecast will faithfully predict your past problems. If your data is spread across several systems, our weekend data audit helps you find out what you have first.

Method 1: Same Period Last Year

The simplest forecast is "this week will look like the same week last year." Forecasters call it a seasonal naive forecast, and it is surprisingly hard to beat for businesses with strong seasons. It captures holidays, summer peaks and back-to-school without any modeling.

Its weakness is that it ignores growth. If the business is 15 percent bigger than a year ago, last year's numbers will be too low every week. That leads to the next step.

Method 2: A Moving Baseline with a Seasonal Index

Split demand into two parts: the level of the business right now, and the seasonal shape of the year.

1. Baseline. Take the average of the last 8 to 13 weeks. This moving average tracks where the business is today and smooths out noisy weeks.
2. Seasonal index. For each week or month of the year, divide its historical sales by the yearly average. A value of 1.3 means that period runs 30 percent above a typical one.
3. Forecast. Remove the season from the recent weeks so the baseline is a typical level, then multiply that level by the index for each future period.

Here is a made-up example to show the arithmetic. Say a shop averaged 200 units a week over the last quarter, and that quarter normally runs at 0.8 of a typical week. The deseasonalized baseline is 250. If December has an index of 1.6, the December forecast is 400 units a week. All of this fits in a spreadsheet, and everyone on the team can follow the logic.

Exponential smoothing is the more refined version of the same idea: recent weeks count more than older ones. Excel's FORECAST.ETS function and most BI tools have it built in.

Method 3: Simple Regression

Regression helps when demand depends on things you know in advance. Typical drivers are price changes, planned promotions, the number of days in the month, holidays that move around the calendar, or the weather for businesses that live outside.

A linear regression with a trend, a few seasonal terms and two or three drivers is enough for many businesses. It also answers questions a moving average cannot, such as roughly how much a 10 percent discount moved volume in past promotions. Keep it small. Each extra variable needs enough history behind it, and a model with twenty inputs fitted on two years of weekly data will memorize the past instead of learning from it.

How to Tell Whether a Forecast Is Any Good

Never judge a forecast by how well it fits the past. Judge it on data it has not seen.

• Hold out recent data. Build the forecast as if it were three or six months ago, then compare it to what actually happened.
• Compare against the naive version. If your regression does not beat "same week last year," use the simpler method.
• Measure error in units people feel. Average error in units or dollars per week means more to a buyer than a percentage score.
• Track it after launch. Put forecast and actual side by side in your weekly report. Drift shows up quickly when people see it every Monday.

Always give a range, not just a number. A forecast of 400 units with a realistic range of 330 to 470 leads to better ordering decisions than 400 on its own.

The Honest Limits

These methods assume the future looks like the past. They struggle when:

• A product is new and has no history
• Sales are low and lumpy, like a few units a month of a spare part
• Something structural changed: a new location, a new channel, a big competitor, a price overhaul
• Your sales history hides lost demand, because you sold out and the orders never happened

In those cases a simple forecast is still a starting point, but you will need judgment layered on top: analog products, planned launches, or input from the people closest to customers.

When to Call in More

It is worth bringing in data science help when:

• You need forecasts for hundreds or thousands of products and locations, automated every week
• Forecast errors are expensive, through spoiled stock, missed sales or overtime
• You have reliable external drivers, such as weather, traffic or bookings, that a model can learn from
• The simple methods are in place and measured, and you can show where they fall short

That last point matters. A data scientist working on top of a clean history and a measured baseline delivers far more than one starting from raw exports. On Figment Forge, our pro bono Power BI dashboard on a real retailer's anonymized data, the forecasting layer was built in Python on historical sales and seasonal patterns, validated against hold-out periods and fed straight into the dashboard. We wrote up how it fits together in the Figment Forge build.

The Bottom Line

You do not need a data science team to forecast demand. You need a clean sales history, a method you can explain, a habit of checking it against reality and a clear decision it feeds. Start with last year plus a moving baseline. Add regression when you have drivers worth modeling. Bring in more when the stakes and the scale justify it.

If you would like help setting that up, or want to know whether your data supports more than the basics, our predictive analytics service starts with the simplest forecast that works and builds from there. Book a free consultation and bring your sales history.