Move the sliders to see how marketing and freight decisions compound into seller growth, output, and revenue over time.
This demonstrates how a forecast model can be designed, built, and deployed to give business leaders concrete insights into past performance, expected outcomes, and how decisions made today affect future results.
A five-equation econometric model of Brazilian e-commerce. Adjust marketing spend and freight subsidies to see how they compound into seller growth, order volume, and revenue over time.
See the data on Kaggle.
How the model works: Marketing investment generates leads that convert into new sellers. The active seller base (net of monthly churn) and product pricing determine how many units sell. Output multiplied by price gives revenue. Freight subsidies reduce the effective shipping cost buyers face, increasing order volume, but the subsidy cost scales directly with each additional unit sold.
Let c index product categories (c = 1, …, 64) and t index months (Feb 2017 – Aug 2018). Define L̃j,c,t = ωj,c Lj,t as the category-weighted MQL count for channel group j ∈ {ps, soc, oth}, where ωj,c is the historical conversion share of channel j in category c and Lj,t is platform-wide lead volume. Define Fc,teff = (1 − τ) Fc,t as effective freight cost net of platform subsidy rate τ ∈ [0, 1]. In each estimated equation, αc denotes category fixed effects, μt month fixed effects, and εc,t the idiosyncratic error.
| Parameter | Eq | Estimate | Interpretation |
|---|---|---|---|
| βps | (1) | 0.122** | +1 weighted MQL → +0.12 new sellers |
| βsoc | (1) | 0.093* | +1 weighted MQL → +0.09 new sellers |
| γ | (1) | 0.436*** | Persistence in new-seller inflows |
| βS | (3) | 5.98*** | +1 seller → +6.0 units/month |
| βF | (3) | −0.224* | +R$1 freight → −0.22 units |
Equations (1) and (3) estimated by within-group (fixed effects) OLS on a balanced panel of 64 categories × 19 months. Eq (1): R2within = 0.162. Eq (3): R2within = 0.673. Exit rates δc in (2) estimated as category-specific monthly attrition rates. Identities (4) and (5) are accounting relationships, not estimated. Source: Olist dataset (Kaggle).
Olist was in a growth phase through the end of the dataset, investing heavily in marketing and subsidies. The model’s most striking insight is structural: at those spend levels, the seller base isn’t sustainable. Estimated monthly exit rates average 33% across categories, meaning roughly 400 of the marketplace’s 1,263 active sellers churn each month. Baseline new-seller inflow (driven by the last observed marketing spend) replaces only about 130. Without increased marketing spend, the model projects seller stock declining to a new equilibrium around 500–600 sellers within 12 months. Try it out: adjust the marketing sliders above and watch the forecast.
Of the channels the marketplace controls, paid search has the largest and most significant effect on new-seller acquisition (β = 0.122, p < 0.01), followed by social media (β = 0.093, p < 0.05). Other channels (organic search, referral, email, and display) are bundled and show no significant aggregate effect. While seven marketing channels exist, two drive actionable growth. Try scaling Paid Search to 2x or 3x above and compare the impact.
Reducing effective shipping costs increases order volume (βF = −0.224, p < 0.10): a 30% freight subsidy adds roughly 2 additional units per category per month. But the cost is borne per-unit (subsidy scales 1:1 with order volume). At 30% subsidy across all 61 categories over 12 months, cumulative cost can exceed R$500K while the incremental revenue gain is modest. The freight lever is real but expensive relative to marketing.
This model was estimated on public data ending August 2018. What happened next is consistent with the model’s core prediction. Olist raised $46M in Series C funding from SoftBank in 2019, followed by $80M in Series D led by Goldman Sachs in early 2021, and then $186M in Series E later that year (achieving a $1.5 billion valuation) [1, 2]. They poured capital into the growth engine and pivoted the business model: launching Olist Shops in 2020, acquiring logistics and commerce-tool companies, and expanding into fulfillment and financial services [2]. Even with that capital, the company later cut staff, citing the need to “grow efficiently” [3]. The treadmill economics the model identifies appear to have shaped the company’s actual strategic trajectory.
The interactive tool is exploratory. The strategy deck distills it into a decision artifact: three preconfigured scenarios with strategic framing, comparative scorecards, and a recommendation. This is the kind of synthesized deliverable that emerges from applying the framework to a real engagement.
Much of the work in this project was in choosing relationships that reflect how the marketplace actually operates, not in tuning for predictive accuracy. This is a methodology demonstration built on public Olist data through August 2018, intended to show how specification-first forecasting enables scenario planning, not to predict Olist’s actual future.
Coefficients are estimated from within-category variation using panel fixed effects. Signs and magnitudes are identified from the data.
Several modeling choices reflect the public-data context: price is treated as observed rather than as a lever (without demand-side structure, treating price as a control variable yields degenerate optimization), channels are grouped into three categories because most have insufficient lead volume to identify separately, and cross-category substitution is excluded.
Applied to real organizational data, the framework would expand to include actual cost structures, real marketing spend rather than lead-count proxies, longer time series, finer lever resolution, and the organizational machinery (sponsorship, success criteria, workflow integration) that makes forecasting drive decisions rather than just describe them.