A dynamic pricing model should use data that describes the market: your own price and sales history, stock levels, costs, competitor prices, and calendar effects such as seasons and campaigns. It should not use data that describes the individual shopper. Data Department, an applied machine learning consultancy in Istanbul, builds dynamic pricing systems for retailers and e-commerce companies on this principle.
Which data describes the market, and which describes the person?
Every pricing input belongs to one of two groups. Market data explains how demand for a product responds to price, time and competition. Personal data explains who is looking at the product. The first group supports a commercial decision you can explain to a customer, a regulator or your own board. The second group increasingly turns pricing into a legal and reputational question.
| Input | What it tells the model | Group | Use it? |
|---|---|---|---|
| Own price and sales history | How demand reacts to price changes (price elasticity) | Market | Yes, the core input |
| Stock level and incoming stock | Whether to protect or clear inventory | Market | Yes |
| Product cost and margin rules | The price floor and the real profit of each sale | Market | Yes |
| Competitor prices | Where the customer's alternative sits | Market | Yes, where reliably collected |
| Calendar, season, campaigns, holidays | Demand shifts that are not caused by price | Market | Yes |
| Product attributes and substitutes | Which products compete with each other | Market | Yes |
| Sales channel (website, marketplace, store) | Different demand and cost per channel | Market | Yes, with consistent rules |
| Browsing history, precise location, demographics | What this particular shopper might accept | Person | No |
Which inputs matter most for a first model?
For most retailers the first model needs only four things: price history, sales history, stock and cost. Price and sales history at SKU (stock-keeping unit) and day level let the model estimate price elasticity. Stock tells it when a lower price is worth it to clear a product. Cost tells it when a lower price destroys margin.
Competitor prices and calendar effects usually come next. They help the model separate the effect of your price from everything else that moved demand that week. Without them, a model may credit a price cut for a sales rise that a public holiday caused.
Why price on the market, not on the person?
Pricing on personal data has drawn regulatory attention. In January 2025 the US Federal Trade Commission (FTC) published initial findings from its surveillance pricing study, describing how intermediaries can use details such as location, demographics and browsing patterns to target individual consumers with different prices. In Türkiye and the European Union, personal data protection law already sets strict conditions on how such data may be processed.
There is also a practical argument. A model that prices on market conditions is easier to test, easier to explain to the commercial team and easier to defend when a customer asks why a price changed. "Stock was high and the season was ending" is an answer. "Our model thought you would pay more" is not.
Dynamic pricing does not mean higher prices. A well-run model lowers prices as often as it raises them, because clearing slow stock and matching competitors are also profitable decisions.
Which data quality problems cause the most trouble?
Pricing models fail more often on data history than on algorithms. Check these before the project starts:
- Price history is not stored. Many systems keep only today's price. Without past prices, the model cannot learn how demand responds to price.
- Promotions are not flagged. A discount recorded as a normal price change makes elasticity look larger than it is.
- Stockouts hide demand. Days with zero stock show zero sales, not zero demand. Those days must be marked, or the model learns the wrong lesson.
- Costs are out of date. In a high-inflation environment, a margin calculated on last quarter's cost can be badly wrong.
- Products change codes. New colours, sizes or packaging often receive new codes, which breaks the link to their sales history.
What does this look like in practice?
In dynamic pricing, we measure success by profitability, not revenue or model error. For an e-commerce retailer, this approach raised profitability by about 4% (pricing case study). The lesson we take from pricing work in general is that profitability, guardrails agreed with the commercial team and a fair baseline matter as much as the model. Our article on why forecast accuracy is not enough explains the same logic for forecasting.
A five-input test for your pricing model
A practical check for any pricing model you already run, or plan to run:
- List the model's five most influential inputs.
- Mark each one as describing the market or the person.
- Ask whether you would be comfortable printing those inputs next to the price.
- Check that each input has reliable history for at least one full season.
- Confirm that the model is judged on profit or margin, not on revenue alone.
If an input fails the third step, remove it before it becomes a legal or reputational problem. A free 30-minute AI (artificial intelligence) opportunity review is a reasonable place to run this test with someone outside the team. Our dynamic pricing page describes how we build these systems.
Frequently asked questions
Is dynamic pricing the same as personalised pricing?
No. Dynamic pricing changes the price of a product for everyone based on market conditions such as stock, season and competition. Personalised pricing charges different people different prices for the same product based on who they are.
How much price history does a dynamic pricing model need?
As a rule, at least one full seasonal cycle, usually a year, with enough price changes to observe how demand responds. Products whose price never changed give the model little to learn from.
Do we need competitor prices to start?
Not always. A first model can start with your own price, sales, stock and cost data. Competitor prices become important in categories where customers compare prices easily.
Can customer segments be used at all?
Segment-level information, such as channel or region, can describe market conditions when it is applied to everyone in that segment equally. Check any use of customer data with your legal team before it enters a pricing model.