Solution / Pricing analytics
Price decisions need evidence, not a model’s opinion.
Pricing analytics connects price, promotion, cost and demand data so a commercial team can test and defend a pricing decision. It is decision support rather than automatic pricing: Quorlytic defines the decision, builds the elasticity or uplift analysis behind it, sets guardrails for margin and service, and measures the outcome against the rule the business used before.
By QuorlyticUpdated
Start from the price decision someone has to make — list, discount, promotion or renewal — then assemble the evidence, the guardrails and the measurement that decision needs. Historical analysis describes past decisions; a controlled test is what supports a new one.
Problem, approach and outcome.
| Problem | Pricing is not one decision. A list-price change, a discount band, a promotion calendar, a renewal offer and a clearance markdown each need different evidence, different guardrails and a different cadence. Write the decision, the owner, the margin or service constraint and the review window before any model is built, because the decision determines the grain, the horizon and what counts as a good answer. |
|---|---|
| Approach | Elasticity estimated from historical sales describes what happened under the prices the business already charged, including the discounts competitors forced and the stockouts that suppressed demand. That is useful context and weak evidence for a new price. Where a controlled test is possible, it answers the question directly; where it is not, state the assumptions explicitly and keep the recommendation bounded. |
| Typical outcome | Pricing decision, guardrail and measurement brief |
The approach
- 01
Name the pricing decision first
Pricing is not one decision. A list-price change, a discount band, a promotion calendar, a renewal offer and a clearance markdown each need different evidence, different guardrails and a different cadence. Write the decision, the owner, the margin or service constraint and the review window before any model is built, because the decision determines the grain, the horizon and what counts as a good answer.
- Decision, owner, cadence and review window stated
- Guardrails agreed for margin, service level and customer impact
- Cost basis agreed before any price analysis
- 02
Separate observation from intervention
Elasticity estimated from historical sales describes what happened under the prices the business already charged, including the discounts competitors forced and the stockouts that suppressed demand. That is useful context and weak evidence for a new price. Where a controlled test is possible, it answers the question directly; where it is not, state the assumptions explicitly and keep the recommendation bounded.
- Historical elasticity labelled as observation, not prediction
- Confounders named: stockouts, promotions, competitor moves, seasonality
- Assumptions and their limits written down with the analysis
- 03
Segment before you generalise
An average price response hides the segments that matter: new against returning customers, contract against transactional, high-against low-volume accounts, and regions with different competitive intensity. Segment-level analysis usually changes the decision more than a better model does, and it shows the business where a uniform price change would be regressive or commercially damaging.
- 04
Promotion effects are not price effects
Promotions change display, timing, availability and customer composition at the same time as price, so a promotion lift is not an elasticity estimate. Model the promotion as its own decision with its own baseline, measure incrementality against a comparable period or holdout, and record the margin cost of the volume it produced.
- 05
What this work should not promise
Pricing analytics does not produce an automatically optimal price, and this page does not claim one. Models are uncertain where history is thin, competitive response is unknown, or the cost basis is contested. The deliverable is a defensible range, the assumptions behind it, the guardrails that bound it and the measurement that shows whether it worked — with human judgement retained for the decision itself.
- No promise of automatic or continuously optimal pricing
- Uncertainty shown rather than a single confident number
- Decision authority stays with the commercial owner
Matching the evidence to the pricing decision
| Dimension | Evidence it needs | Where the evidence stops |
|---|---|---|
| List price change | Segment-level response analysis and, where feasible, a bounded test. | Competitor response is not observable in your own history. |
| Discount band | Win rate and margin by segment and deal size. | Sales behaviour changes as soon as the band changes. |
| Promotion | A holdout or comparable-period incrementality estimate with its margin cost. | Display, timing and availability move at the same time as price. |
| Renewal offer | Churn sensitivity by tenure and usage, plus margin by cohort. | Small samples in the highest-value segments. |
| Clearance markdown | Sell-through curves and remaining-life analysis. | Demand is truncated by stock-outs and the end of the season. |
What the work produces
- Pricing decision, guardrail and measurement brief
- Elasticity and segment analysis with stated limits
- Controlled test design where a test is feasible
- Margin and outcome review pack
Delivery stages
- 01Name the decision
- 02Assemble the evidence
- 03Test or bound it
- 04Measure and review
Identify where a proposed price has no comparable evidence.
A fitted price-response curve can produce a number for a price the business has never charged in comparable conditions. That numerical output is extrapolation, not additional evidence. In this hypothetical discount analysis, the useful deliverable maps which customer/product conditions support a comparison and which proposed changes remain scenarios requiring a different decision basis.
Scope and assumptions: The example is decision support for a permitted commercial analysis. It supplies no recommended price, elasticity estimate or profit forecast. Historical comparisons need an explicit causal rationale, adequate support and a suitable cost definition; automatic pricing and competitive coordination are outside this example.
| Dimension | Proposed comparison | Evidence boundary and next decision |
|---|---|---|
| Price actually paid | Define the effective price after the discounts and credits relevant to the decision. Keep tax, currency, pack size and contract conditions consistent. Decide whether returns change the outcome window or the cost calculation. | Reconcile a representative invoice and adjustment trail before estimating a response. A list-price field and a realised unit price answer different questions and cannot be substituted silently. |
| Comparable conditions | Inspect the products, customer groups, periods and availability conditions in each price range. A discount used only during clearance is mixed with a different stock position and customer population. | Record the proposed causal assumptions and pre-decision confounders. Adjustment cannot recover comparisons absent from the data, and including more variables does not automatically establish an identifiable price effect. |
| Support and overlap | Identify segments observed under materially different prices in comparable conditions. If premium accounts always received one contract rate, a model's alternative-rate estimate for that group may depend almost entirely on extrapolation. | Keep a support map and the excluded population with the estimate. DoWhy provides an overlap assessment for certain backdoor-adjustment estimators; its documented scope does not make every continuous-price model or arbitrary comparison valid. |
| Proposed price range | Separate interpolation within supported conditions from a scenario beyond observed prices, product mix or capacity. Include a no-change option and operational guardrails for margin, service and customer impact. | Label an unsupported range as a scenario assumption. If a controlled test is suitable, define its eligible population, randomisation unit, decision window and stop conditions before using its results to make a broader claim. |
| Portfolio outcome | Check whether customers can substitute another product or delay a purchase. A promoted item's additional units may partly replace other margin rather than add to total portfolio contribution. | Retain the product group and outcome horizon needed to observe substitution, returns and delayed purchases. Report the supported population and uncertainty alongside any result rather than extending a narrow comparison to the entire catalogue. |
At narrow widths, scroll the worksheet to read every column. With the table focused, use the arrow keys to scroll.
Worked decision: a model suggests a deeper discount for a segment whose historical discounts occurred only in clearance periods. Keep the output as a scenario, document the missing comparable conditions and examine whether a bounded test or a narrower supported comparison is feasible. A smooth response curve does not fill that evidence gap. No price change or margin improvement is asserted by this example.
- Avoid adjusting indiscriminately for variables created by the price decision itself. Choose the causal structure and timing before choosing controls; the modelling library cannot decide those business assumptions for you.
- Keep feasibility separate from expected response. Contract terms, capacity, inventory and customer obligations can rule out a proposed action even when an analytical comparison is statistically credible.
When price, promotion and planning share inputs, connect the decision to demand forecasting and planning analytics while keeping scenario assumptions distinct from known future plans.
Official references for this example:
Questions
Questions worth asking.
Can you tell us the optimal price?
No, and any answer offered without your cost basis, competitive context and a test would be a guess with a model attached. What we can do is quantify how demand has responded, identify where the data cannot support a conclusion, design a bounded test, and measure the margin outcome against the rule you use today.
Do we need machine learning for pricing?
Not always. Segment-level analysis, a controlled test or a simple decision rule often answers the question with less to maintain. Model complexity earns its place when the decision needs a distinction that rules cannot express — for example heterogeneous response across many products, regions or customer types — and when someone can act on that distinction.
Your next data decision
Start with a clear problem.
Build what comes next.
Bring the business question.
We'll help shape the engineering brief.