Siam Sports: Promotion Profitability and Regression Traps
Siam Sports operates 62 sporting-goods stores. Marketing spend is THB 10 million a month: THB 3 million TV, THB 1 million radio, THB 2 million digital and THB 4 million price promotions. A consultant regressed monthly store sales on spend over 36 months: promotion coefficient 4.2 (p<0.01), digital 1.8 (p=0.21), TV 0.6 (p=0.04); R²=0.71 and Durbin–Watson 0.9. Promotions are deeper in December and Songkran. A four-week test in 20 stores offered 20% off all footwear; unit sales rose 35% from a baseline of 1,000 units per store per month at THB 100 average price; margin fell from 38% to 22%. In the four weeks after the test, units ran 12% below the pre-test baseline. The CMO wants to shift all promotion money into digital based on the regression. Finance asks whether the footwear test proved promotions work.
1. Evaluate the profitability of the 20% footwear promotion, including the post-promotion dip. Would you repeat it? Show calculations.
Read the worked answer
Baseline per store per month. 1,000 units × THB 100 = THB 100,000 revenue; gross profit at 38% = THB 38,000.
Promotion. 1,350 units × THB 80 = THB 108,000 revenue; gross profit at 22% = THB 23,760. Incremental gross profit = 23,760 − 38,000 = −THB 14,240 per store.
Post-promotion dip. 12% × 1,000 = 120 units lost at THB 38 contribution = −THB 4,560. Total effect ≈ −THB 18,800 per store, or −THB 376,000 across 20 stores.
To break even, the promotion needed about 495 units of full-price-equivalent contribution (18,800 ÷ 38). It delivered 350 incremental units, so it would have needed roughly 145% more incremental volume.
Do not repeat a 20% blanket discount at a 38% margin. The discount is too deep. Test 10–12% off a narrower SKU set, require a minimum basket, and measure whether promo buyers return at full price. Only a large share of genuinely new, retained customers would justify this economics.
2. Critique the regression as a basis for reallocating budget. Address at least three threats and state what analysis you would run instead.
Read the worked answer
Threat 1: endogeneity. Promotions are timed to weak months and holidays, so the coefficient partly captures seasonality and price sensitivity, not causal lift.
Threat 2: autocorrelation. Durbin–Watson 0.9 indicates positive serial correlation. Standard errors are understated, p-values are too optimistic and the digital coefficient may appear insignificant only because the model is misspecified.
Threat 3: multicollinearity. Media channels and promotions move together, inflating variance and making individual coefficients unstable. Digital's p=0.21 is not evidence that digital does nothing.
Threat 4: aggregation and omitted variables. Monthly store sales hide store heterogeneity, saturation and competitor actions.
Instead of reallocating on this regression, build a weekly marketing-mix model with adstock and saturation terms, store fixed effects and seasonality controls, and validate it with planned experiments. Do not cut TV merely because p=0.04; statistical significance is not economic significance, and TV may build brand effects that appear with a lag. The footwear test is more informative about promotions than the regression, and it says the promotion destroyed margin.
3. Design a geo holdout to estimate the true incremental return of digital versus promotion, including store matching, duration, and decision rules.
Read the worked answer
Select 40 stores and match them into 20 pairs on prior sales, store format, region and digital exposure. Randomize one store in each pair to treatment: increase digital spend by 50% for 12 weeks. Hold promotion constant in both arms so digital is isolated.
Primary KPI: incremental revenue per store per week. Secondary KPIs: gross profit after media, new-customer count and digital-attributed conversions. Pre-register a minimum detectable effect, for example a 3% sales lift, and check whether 20 pairs over 12 weeks can detect it. If not, extend the duration or use more stores.
Decision rule: scale digital only if the 95% confidence interval on incremental gross profit excludes zero and the return exceeds the cost of capital. If the interval includes zero, keep the current mix and run a second, longer test rather than reallocating on the regression. Report both intention-to-treat and per-protocol results, and do not change promotion depth mid-test.