Skip to main content
AI // Forecasting

Forecasts that go straight into the purchase order.

Weekly demand forecasts by product and location, trained on your history and fed into replenishment rules, so buyers start from a smart suggestion instead of a gut feeling.

inference.tracelive

{

"summary"
: "Forecast, compare with actuals, adjust. A forecast nobody uses is just a chart.",
"granularity"
: "SKU × location × week",
"inputs"
: "Sales · seasonality · promotions",
"output"
: "Replenishment suggestions",
"reviewed_by"
: "a human"

}

From history to purchase order

  1. Step 01

    Data audit

    Check history length, stockouts that hide true demand, and product changes.

  2. Step 02

    Baseline

    Start with a simple statistical forecast to beat.

  3. Step 03

    Model

    Add machine learning features like seasonality, promotions, and price where they improve accuracy.

  4. Step 04

    Embed

    Feed forecasts into Odoo reorder rules and purchase suggestions.

  5. Step 05

    Monitor

    Track accuracy per category and retrain on schedule.

Beat the baseline or don't ship

Fancy models must outperform a simple moving average on your own data. If they don't, the simple model wins and costs less to run.

Why spreadsheets stop working

Most purchasing still runs on last year's numbers, a spreadsheet, and a buyer's instinct. That works for a few hundred products. Across thousands of SKUs, several locations, and seasonal swings, it leads to the usual pattern: too much of what doesn't sell, not enough of what does, and cash tied up in the wrong stock.

Statistical and machine learning forecasts help by looking at every item's history at once, separating trend, seasonality, and one-off spikes, and producing a forecast with a range rather than a single guess. The value comes when that forecast drives reorder points and purchase suggestions automatically, so buyers spend their time on exceptions rather than routine.

Forecasting is only as good as the history behind it, so every project starts with an honest look at the data: stockouts that hid real demand, product codes that changed, promotions nobody recorded.

Signals a forecast can use

SignalExampleWhen it helps
Sales historyWeekly units by SKU and locationAlways: the foundation
SeasonalityHoliday peaks, summer dipsProducts with repeating patterns
Promotions and priceDiscount weeks, price changesRetail and e-commerce catalogs
Calendar effectsPublic holidays, month endsB2B and retail alike
Stockout flagsDays with zero stockCorrecting history that understates demand
Product hierarchyCategory, brand, sizeNew and slow-moving items

What you get

Forecasts with ranges

A likely value plus a low and high band, so safety stock reflects real uncertainty.

Replenishment hooks

Forecasts update reorder rules and purchase suggestions in Odoo.

Exception views

Items where the forecast changed sharply or accuracy dropped are surfaced for review.

Accuracy tracking

Forecast versus actual by category and period, so trust is earned with numbers.

Delivery plan

  1. Step 01

    Data audit

    Week 1–2

    Assess history, gaps, stockouts, and product changes; agree the forecast level.

  2. Step 02

    Baselines

    Week 2–3

    Build simple benchmarks that any model must beat.

  3. Step 03

    Models and backtests

    Week 3–5

    Train candidate models and test them on past periods they never saw.

  4. Step 04

    Integration

    Week 5–7

    Connect forecasts to Odoo replenishment and build the review screens.

  5. Step 05

    Run and refine

    Ongoing

    Monitor accuracy, retrain on schedule, and tune categories that underperform.

Forecasting myths

Myth: Machine learning always beats simple methods.

Reality: Often it doesn't, especially with short or noisy history. That's why every model must beat a baseline on your data.

Myth: A forecast should be one number.

Reality: A single number hides risk. Ranges set safety stock sensibly and show where uncertainty is high.

Myth: Once built, a forecast runs forever.

Reality: Demand patterns shift. Accuracy is monitored and models are retrained on a schedule.

Common pitfalls we help you avoid

The first pitfall is trusting history that hides real demand. When an item was out of stock, recorded sales understate what customers wanted, so we flag stockout periods and adjust before training.

The second is forecasting at the wrong level. Very fine detail can be too noisy to predict, while very coarse forecasts can't drive purchasing, so we agree the level with buyers before building anything.

The third is a forecast that lives in a separate report. If it doesn't feed reorder rules and purchase suggestions, buyers keep their spreadsheets, and the investment never pays back.

What we measure

MeasureWhy it matters
Forecast error by categoryWhere forecasts are reliable and where they aren't
BiasWhether forecasts are consistently too high or too low
Stockout and overstock daysThe business effect buyers actually care about
Override rateHow often buyers change suggestions, and whether it helps

Who does what

We handle

  • Data preparation, models, and backtesting
  • Odoo integration and review screens
  • Monitoring and scheduled retraining

Your team handles

  • Sales history and product data access
  • Planned promotions and known events
  • Approving purchase suggestions

A worked example

Take a retailer with a few thousand SKUs across several stores and a central warehouse. Buyers review stock weekly from a spreadsheet and reorder by feel, so popular items run out before holidays while slow lines pile up. A weekly forecast by SKU and location, with ranges, feeds reorder points in Odoo.

Buyers start each week from purchase suggestions instead of a blank sheet, and spend their time on the items flagged as unusual. Forecast accuracy, stockouts, and overstock are tracked side by side, so the team can see whether the change is working.

Before you start

  • At least a year of sales history, ideally more
  • Records of stockouts and promotions
  • Supplier lead times in the ERP
  • A buyer who will review suggestions weekly

What success looks like

Success is buyers who start from the forecast rather than a spreadsheet, fewer emergency orders, and less slow-moving stock, with accuracy reported openly so everyone knows where the forecast can be trusted and where judgment still matters.

What makes demand forecastable

Forecasting: Principles and Practice, a widely used open textbook, lists four factors that determine how predictable something is: how well we understand what drives it, how much data is available, how similar the future is to the past, and whether the forecast itself affects the outcomeSource [1]. Promotions and stockouts often break the last two, which is why we model them explicitly.

In practice we start with a simple baseline, such as last year's pattern adjusted for trend, and only keep a more complex model if it beats the baseline on your history. Forecasts then feed reorder suggestions inside Odoo, with planners able to override them and the reason recorded.

Frequently asked questions

How much history do we need?
More is better, ideally covering a few seasonal cycles. The data audit tells you what's realistic.
Can it handle new products?
New items borrow patterns from similar products until they have their own history.
Do buyers lose control?
No. Forecasts produce suggestions; buyers approve or adjust them.
Which tools do you use?
Python forecasting and machine learning libraries, running as a service connected to Odoo.
How is accuracy reported?
Error metrics by category and period, visible in a dashboard next to actual sales.
What forecast horizon do you use?
Usually weekly forecasts a few months ahead, matched to your supplier lead times and planning cycle.
Can promotions be planned in?
Yes. Planned promotions can be entered so their expected uplift is included in the forecast.
Does it work for B2B with lumpy demand?
Lumpy demand is harder to forecast. We use methods designed for intermittent demand and set expectations honestly.
Where does the forecasting run?
In a separate service on a schedule, writing results back to Odoo through its API.
Can buyers override the forecast?
Always. Overrides are recorded, which also shows whether manual adjustments improve or hurt accuracy.
Can it forecast at the warehouse level?
Yes, by SKU and location, provided each location has enough history or can borrow patterns from similar ones.
Does it account for supplier lead times?
Yes. Lead times and minimum order quantities are used when turning forecasts into purchase suggestions.
How often is the forecast refreshed?
Usually weekly, aligned with the purchasing cycle, with more frequent refreshes for fast-moving items if useful.
What if our data is messy?
Most sales data is. The audit finds the issues, such as stockouts, renamed products, and missing promotions, and we fix or work around them before modeling.
Can we see forecasts in Odoo directly?
Yes. Forecasts and suggested quantities appear on product and replenishment screens buyers already use.
Why do some products forecast badly?
Usually limited data, changing conditions, or drivers like promotions that aren't in the model, all factors known to limit predictabilitySource [1].
Will forecasts replace our planners?
No. They give planners a better starting point with the reasons attached, and planners keep the final decision.