Devaneios

Joguei meus dados de corrida numa IA só pra brincar. Não esperava sair com uma reflexão sobre análise de dados. (sim, eu sei que isso soa exatamente como todo post de LinkedIn, mas poh, o exemplo que saiu é bom demais pra não compartilhar)

Exportei o histórico do Garmin e pedi pra analisar meu pace ao longo do tempo. O gráfico ficou bonito. Curva estável, evolução sutil.

Aí a IA apontou um ponto fora da curva. Dia 15/03. Pace de 6:50/km, o melhor da série. FC de 177 bpm, a mais alta de todas. Diagnóstico: possível outlier, pode distorcer a análise.

Ela estava certa, com as informações que tinha.

O que a IA não sabia: aquele dia era o teste de 5 km. O objetivo de meses treinando 3x por semana desde setembro. Meu tempo anterior era 34:50. Naquele dia fui pra 34:12.

Não era um erro no dado. Era o dado mais importante de todos.

Fica mais fácil de ver o problema quando você substitui a IA por um analista que só recebeu o gráfico, sem nunca ter conversado com a operação, sem saber o que aquele pico representava. A conclusão seria a mesma: remove o outlier, entrega o relatório limpo, conta a história errada.

Dado sem contexto é ruído. E às vezes o outlier é o resultado.

A metric doesn't have to be wrong to conceal a problem.

At a large manufacturer I worked for, the import team relied on one number: landed cost factor, the total cost to import a product, divided by the product's value. It's the metric supply chain uses to determine whether freight contracts are effective.

For a period in 2024, international freight rates kept increasing. Some products, the ones with the least predictable demand, absorbed the full impact. But across the company, the average remained stable. Contracts, negotiation, mix: the factor stayed healthy. Nothing alarming on the dashboard.

Meanwhile, the business units, my internal clients, were watching margin erode, and were quick to blame the obvious suspect: "freight's too expensive, you're just not negotiating hard enough."

I isolated the outliers and reviewed them with the cost team. Freight had barely moved the factor. What had moved, substantially, throughout that year, was the exchange rate.

Here's the nuance that's easy to overlook. That factor is a dollar-over-dollar ratio: FOB, freight, everything on the import side is priced in dollars. When the exchange rate surges, the ratio remains unaffected, because both sides of the division move together. The metric is constructed to be insensitive to currency movement, by design.

What it also fails to capture is the four months between placing an order and that product's impact materializing in a customer's margin: transit, factory adjustment, warehouse time, the sale itself. Prices hadn't been repriced for that window's exchange rate. By the time the impact was visible on the sales side, the currency movement responsible for it was five months old, already forgotten.

The freight number wasn't inaccurate. It just wasn't constructed to detect what was actually happening.