Price elasticity of demand: how to estimate it from data without being fooled
How to obtain price elasticity of demand from sales history: log-log regression, point and arc elasticity, cross elasticity, the optimal price and the main traps (seasonality, promotions, cannibalization). The result decides how far a price can be changed, what volume response to expect, and when the estimate cannot be trusted.
Why this document
The pricing lead and the commercial director read this before changing a price or running a promotion. It matters because prices are moved on how demand will react, and a naive estimate from history usually lies: a low price coincides with a weak season, so the relationship looks weaker than it is. Without controlling for season and promotion it is easy to underestimate sensitivity and cut a price for nothing, or raise one for nothing. An estimate with controls and an interval shows what can be relied on.
At a glance
- What is this document and why is it needed?
- How to obtain price elasticity of demand from sales history: log-log regression, point and arc elasticity, cross elasticity, the optimal price and the main traps (seasonality, promotions, cannibalization). The result decides how far a price can be changed, what volume response to expect, and when the estimate cannot be trusted.
- Who owns the document and how often is it rebuilt?
- Pricing or commercial analyst; estimated once a season by category and before every major price change; read by the buyer, the product director and the finance lead.
- What mistakes are the most common?
- Elasticity is computed from price alone without a season control → a biased estimate (understated in the example) → add the seasonal index and a promo flag. The estimate is extrapolated beyond the price range → an unrealistic forecast → apply only within the observed range, pilot for the rest. The post-promotion dip is ignored → the effect is overstated → use a window with the post-promotion trough.
For whom
- Brand
- A brand changes the recommended retail price and wholesale terms rarely and for a whole season: elasticity in own channels (D2C, marketplaces) is estimated from price and promotion history, and for wholesale from sell-out response at different price points. An error in elasticity costs a whole season of margin.
- Retail
- A retailer changes prices often (promotions, markdowns) and has many observations by store and week: elasticity is estimated by category and price band with controls for season and promotion, and cross elasticity is used to gauge cannibalization within a category.
What it consists of
What elasticity is and how to read it
- Elasticity E = percentage change in volume per 1% change in price; with E = −2 a 10% price cut gives roughly a 20% volume rise. In absolute value |E| > 1 means elastic demand (revenue rises when price falls), |E| < 1 inelastic.
- Elasticity is not a constant: it depends on the price range, season, item role (basic or fashion) and competitors; an estimate is valid only within the observed price range.
Estimation methods
- Arc: E = [(q₂ − q₁)/((q₁ + q₂)/2)] ÷ [(p₂ − p₁)/((p₁ + p₂)/2)]; for price 4,000 → RUB 3,200 and volume 80 → 132 units it is -2.21, the point (log) estimate is -2.24. The two numbers are close for small changes; for deep changes arc describes an episode more reliably.
- Log-log regression: ln q = a + E·ln p + controls + error; the coefficient on ln p is the elasticity (constant in the model). Controls: seasonal index, promotion and display flag, stock availability, competitor price.
- Cross elasticity E_AB = (% change in volume of B) ÷ (% change in price of A): with a 20% price cut on A and a 6% drop in volume of B it is 0.3: the items substitute for each other and part of A growth is cannibalization of B.
How not to be fooled
- The main trap is price coinciding with season: prices are cut when demand is falling anyway. In the example (52 weeks, true value 1.8) the naive one-variable regression gives 1.1 while with season and promotion controls it is 1.8 ± 0.36: the naive estimate understates sensitivity by 39%.
- A promotion simultaneously cuts price and adds display, mailing and shelf space: without a separate promo flag the advertising effect is credited to price and elasticity is overstated.
- Pull-forward: sales dip after a promotion; elasticity must be computed over a window that includes the post-promotion dip, otherwise the effect is overstated.
- Brand: prices differ across partners and channels and are linked, so elasticity is computed by channel and flows between channels are checked. Retailer: prices are cut non-randomly (on leftovers), so stock and item age must be controlled.
The profit-maximizing price
- With constant elasticity e (absolute, e > 1) and cost c the optimal price is p* = c·e/(e − 1): for c = RUB 2,000 and e = 2.2 that is RUB 3667. This is an upper bound of the reasoning: the real price is limited by positioning, competitors and the price ladder.
- At |e| ≤ 1 the formula is undefined: profit rises with price and only competition and brand perception limit it.
Brand and retailer: elasticity under different conditions
- Brand: price for a brand is the recommended retail and the wholesale price: elasticity in retail channels is estimated on own sales, and for wholesale the partner sell-out response to its price change is used.
- Retailer: price for a retailer is the shelf price: it has many observations by store and week, so a panel regression with store, season and promotion controls is possible.
- Brand: brand prices change rarely and in steps (season, price list review), so there is little variation and the estimate is backed by a price test on part of D2C traffic or on a marketplace.
- Retailer: retailer prices change often but not randomly; to separate the price effect from the effect of leftovers and item age, age and stock are added to the model.
- Brand: a brand computes cross elasticity between its lines and between channels: a price cut in D2C shifts partner sales and this flow enters the effect estimate.
- Retailer: a retailer computes cross elasticity between substitutes within a category and between brands; cannibalization within the category shows whether the model brought growth or merely shifted sales.
- Brand: a brand agrees markdown depth with partners: a too deep brand promotion in D2C devalues partner leftovers, so elasticity is complemented by an estimate of partner reactions.
- Retailer: a retailer chooses markdown depth by elasticity and by the leftover liquidation price: elasticity together with the payoff threshold shows when depth stops working.
- Brand: a brand decides on price a season ahead, so the elasticity interval is turned into revenue scenarios (low, middle, high elasticity).
- Retailer: a retailer revisits the price decision weekly, so the elasticity estimate is updated over a rolling window and used for automatic markdown rules.
Where the data comes from
| Data source | Fields |
|---|---|
| Weekly sales and prices | Volume, price excluding VAT (before and after discount), share of promotional sales by SKU, store and week; at least 40-60 observations per group. |
| Promotion and event calendar | Promotion type, display, mailing, holidays, weather anomalies, stock availability. |
| Competitor prices | Price monitoring of comparable items; without it elasticity mixes with competitor reactions. |
How to build it step by step
- Step 1. Define the estimation group: a category and price band, not a single SKU (too few observations per SKU).
- Brand: Step 2. Take the series by channel (D2C, marketplace) separately and add partner prices of the same model to the controls.
- Retailer: Step 3. Take the series by week and store and add controls for item age and stock to separate a price cut from natural decay.
- Step 4. Extract the seasonal index (see demand forecasting) and a promo flag; reconstruct demand on out-of-stock days.
- Step 5. Plot ln q against ln p and look at the cloud: if price barely varies or varies only with season the estimate will be unreliable.
- Step 6. Estimate the regression of ln q on ln p without and with controls; compare coefficients and get the standard error and interval.
- Step 7. Check plausibility: negative sign, magnitude in reasonable bounds, interval excludes zero; estimates by category should not contradict common sense.
- Step 8. Estimate cross elasticity for pairs of substitutes and include cannibalization in the promotion forecast.
- Step 9. Confirm the estimate by an experiment where possible (a price test on a group of stores or part of traffic) and recompute.
- Step 10. Apply the estimate only within the observed price range; for going beyond it plan a pilot.
Formulas
| Measure | Formulas | Note |
|---|---|---|
| Point elasticity | E = ln(q₂/q₁) / ln(p₂/p₁) (for small changes E ≈ (dq/q)/(dp/p)) | Symmetric with respect to the direction of the change. |
| Arc elasticity | E = [(q₂ − q₁) / ((q₁ + q₂)/2)] / [(p₂ − p₁) / ((p₁ + p₂)/2)] | Describes one episode; does not remove the effect of season. |
| Log-log model | ln q_t = a + E·ln p_t + β₁·ln s_t + β₂·promo_t + ε_t | s_t seasonal index; E is the constant elasticity; OLS. |
| Estimate interval | E ± 1.96·SE(E), SE = √( s²·[(XᵀX)⁻¹]_(22) ), s² = SSE/(n − k) | Assumes independent residuals; with autocorrelation the interval is too narrow. |
| Cross elasticity | E_AB = (Δq_B / q_B) / (Δp_A / p_A) | Positive for substitutes, negative for complements. |
| Optimal price | p* = c · e / (e − 1), e = |E| > 1 | Only with constant elasticity; c is cost excluding VAT. |
| Volume change under a price move | q₂ / q₁ = (p₂ / p₁)^E | For E = −2 a 30% price cut gives a 2.04-fold volume rise. |
Formulas in notation
- OLS estimate
β̂ = (XᵀX)⁻¹ Xᵀ y; SE(β̂_j) = √( σ̂²·[(XᵀX)⁻¹]_{jj} )
Solution of the normal equations and the standard error.- Profit as a function of price
π(p) = (p − c)·q₀·(p/p₀)^{−e}
The profit curve for the chart.
Worked example
Elasticity estimated from 52 weeks: naive, with controls and by category
| Estimate | Elasticity (absolute) | Standard error | True model value |
|---|---|---|---|
| Naive (price only) | 1.1 | 0.19 | 1.8 |
| With season and promo controls | 1.8 | 0.19 | 1.8 |
| Apparel | 1.68 | 0.18 | 1.6 |
| Footwear | 1.36 | 0.17 | 1.3 |
| Accessories | 2.11 | 0.19 | 2.1 |
| Kids | 1.04 | 0.19 | 1.1 |
| Men | 1.36 | 0.18 | 1.4 |
| Women | 1.66 | 0.2 | 1.9 |
Weekly series are generated in code from a known elasticity (demand = level × season × price^−e × promo effect × noise), prices are cut mostly in the weak season. The naive estimate 1.1 is understated; the estimate with controls 1.8 at R² 0.95 is closer to the truth. The data is illustrative: on real data the direction of the bias can be any.
Figures
- points
- regression line: y = 6.323 − 1.102·x
Data as a table
| ln price (price index) | ln weekly volume | |
|---|---|---|
| −0.36 | 6.55 | |
| −0.32 | 6.55 | |
| −0.27 | 6.46 | |
| −0.09 | 6.03 | |
| −0.11 | 6.17 | |
| −0.26 | 6.53 | |
| −0.18 | 6.53 | |
| −0.02 | 6.11 | |
| −0.33 | 6.74 | |
| −0.25 | 6.58 | |
| 0 | 6.15 | |
| −0.13 | 6.48 | |
| −0.01 | 6.23 | |
| −0.01 | 6.34 | |
| −0.23 | 6.81 | |
| −0.29 | 6.93 | |
| −0.01 | 6.44 | |
| −0.01 | 6.37 | |
| −0.3 | 7.08 | |
| −0 | 6.39 | |
| −0.3 | 6.97 | |
| −0.01 | 6.48 | |
| −0.01 | 6.41 | |
| −0.02 | 6.54 | |
| −0.01 | 6.46 | |
| −0.01 | 6.47 | |
| −0.17 | 6.94 | |
| −0.02 | 6.41 | |
| −0.01 | 6.5 | |
| −0.02 | 6.42 | |
| −0.02 | 6.48 | |
| −0.13 | 6.68 | |
| −0.02 | 6.44 | |
| −0 | 6.43 | |
| −0.01 | 6.37 | |
| −0.01 | 6.38 | |
| −0.01 | 6.26 | |
| −0.01 | 6.4 | |
| −0.02 | 6.22 | |
| −0.01 | 6.2 | |
| −0.01 | 6.17 | |
| −0.01 | 6.13 | |
| −0.29 | 6.61 | |
| −0.01 | 6.06 | |
| −0.32 | 6.69 | |
| −0.26 | 6.53 | |
| −0.32 | 6.72 | |
| −0.35 | 6.63 | |
| −0.34 | 6.64 | |
| −0.31 | 6.53 | |
| −0.31 | 6.4 | |
| −0.21 | 6.2 |
Naive estimate: ln volume against ln price. The line through the cloud has a slope weaker than the true one: low prices coincide with the weak season.
- points
- regression line: y = 6.181 − 1.801·x
Data as a table
| ln price (price index) | ln volume net of season and promo | |
|---|---|---|
| −0.36 | 6.83 | |
| −0.32 | 6.81 | |
| −0.27 | 6.7 | |
| −0.09 | 6.33 | |
| −0.11 | 6.43 | |
| −0.26 | 6.66 | |
| −0.18 | 6.61 | |
| −0.02 | 6.25 | |
| −0.33 | 6.75 | |
| −0.25 | 6.54 | |
| 0 | 6.18 | |
| −0.13 | 6.46 | |
| −0.01 | 6.18 | |
| −0.01 | 6.26 | |
| −0.23 | 6.61 | |
| −0.29 | 6.7 | |
| −0.01 | 6.28 | |
| −0.01 | 6.18 | |
| −0.3 | 6.78 | |
| −0 | 6.17 | |
| −0.3 | 6.64 | |
| −0.01 | 6.23 | |
| −0.01 | 6.15 | |
| −0.02 | 6.27 | |
| −0.01 | 6.19 | |
| −0.01 | 6.2 | |
| −0.17 | 6.57 | |
| −0.02 | 6.14 | |
| −0.01 | 6.23 | |
| −0.02 | 6.16 | |
| −0.02 | 6.23 | |
| −0.13 | 6.45 | |
| −0.02 | 6.22 | |
| −0 | 6.22 | |
| −0.01 | 6.19 | |
| −0.01 | 6.22 | |
| −0.01 | 6.12 | |
| −0.01 | 6.29 | |
| −0.02 | 6.14 | |
| −0.01 | 6.15 | |
| −0.01 | 6.15 | |
| −0.01 | 6.15 | |
| −0.29 | 6.58 | |
| −0.01 | 6.15 | |
| −0.32 | 6.74 | |
| −0.26 | 6.62 | |
| −0.32 | 6.85 | |
| −0.35 | 6.79 | |
| −0.34 | 6.84 | |
| −0.31 | 6.77 | |
| −0.31 | 6.67 | |
| −0.21 | 6.48 |
After controlling for season and promotion. The season and promotion contributions are removed from ln volume: the slope is close to the true value of 1.8.
Data as a table
| Measure | Elasticity (absolute) |
|---|---|
| Naive | 1.1 |
| With controls | 1.8 |
| Lower 95% bound | 1.44 |
| Upper 95% bound | 2.16 |
| True value | 1.8 |
Elasticity estimates and the true value. Controls bring the estimate back toward the truth; the interval shows the residual uncertainty.
- Estimate
- Set value
Data as a table
| Measure | Estimate | Set value |
|---|---|---|
| Apparel | 1.68 | 1.6 |
| Footwear | 1.36 | 1.3 |
| Accessories | 2.11 | 2.1 |
| Kids | 1.04 | 1.1 |
| Men | 1.36 | 1.4 |
| Women | 1.66 | 1.9 |
Estimates for six slices and the set values. The estimate with controls recovers the set values within error.
Data as a table
| Measure | Profit, RUB thousand |
|---|---|
| 2 500 | 141 |
| 2 800 | 175 |
| 3 100 | 193 |
| 3 400 | 200 |
| 3 667 | 202 |
| 4 000 | 200 |
| 4 400 | 195 |
| 4 800 | 188 |
| 5 200 | 180 |
| 5 600 | 172 |
Profit as a function of price at e = 2.2. The maximum is near RUB 3,670 (p* = c·e/(e−1)); towards higher prices profit falls more slowly than towards lower ones.
Specifics by category
- Apparel
- Elasticity is higher on seasonal and fashion items and lower on basics; compute by price band and role, not for the whole category; at season end elasticity falls with interest in the item.
- Footwear
- Purchases are pricier and rarer, choice depends on size and feel: price elasticity is moderate while size availability matters strongly; without a stock control it is understated.
- Accessories
- Gift purchases are less price-sensitive at peaks and more so outside them; estimate elasticity separately for the peak and the off-season.
- Kids
- Need-driven purchases (size, school) make demand less elastic on essential items; for fashionable kids items elasticity is closer to women apparel.
- Men
- Demand is steadier and less sensitive to short promotions but sensitive to price level and brand; cross effects among basics are weaker.
- Women
- A wide choice and strong substitutes give high cross elasticity and cannibalization within a category; a promotion on one model shifts sales from neighbors.
Common mistakes
- Elasticity is computed from price alone without a season control → a biased estimate (understated in the example) → add the seasonal index and a promo flag.
- The estimate is extrapolated beyond the price range → an unrealistic forecast → apply only within the observed range, pilot for the rest.
- The post-promotion dip is ignored → the effect is overstated → use a window with the post-promotion trough.
- Cannibalization is not counted → growth of the promoted model is taken for category growth → estimate cross elasticity and look at the whole category.
- The optimal price from the formula is taken as final → the price ladder and positioning are broken → use p* as an upper guide, not a target.
Russian specifics
- Price for elasticity is taken excluding VAT: the 22% rate does not change relative price change but a rate change inside the series creates a false shelf-price jump. Rule: Tax Code Art. 164: Tax Code, Art. 164 (VAT rates).
- The recommended retail price of a brand for partners must not become price coordination between competitors; the exchange of price information with partners is limited by competition-protection requirements. Rule: Law 135-FZ of 26.07.2006: Federal Law of 26.07.2006 No. 135-FZ on protection of competition (rule check not completed).
- Promotional prices are subject to advertising law requirements on conditions and terms; promotion terms come from the calendar that also feeds elasticity. Rule: Law 38-FZ of 13.03.2006: Federal Law of 13.03.2006 No. 38-FZ on advertising (rule check not completed).