Forecasting income is not the same as forecasting weather. Weather models feed on physical measurements: pressure, humidity, temperature gradients. Income models for delivery couriers feed on behavioral patterns: when you work, how long, which platform, which part of the city. The data exists. The challenge is knowing which parts of it actually predict next week's earnings and which parts are noise.
When I started building the forecasting model at Trampay, my expectation was that total hours worked would be the dominant predictor. More time on the road, more deliveries, more income. That is broadly true but it turns out to be a fairly weak predictor in isolation. What actually drives forecast accuracy is a combination of four signal categories that interact in ways that are not immediately obvious.
Signal One: Order Density Within Your Active Windows
Every courier has a pattern of when they are active. Some work every morning. Some work evenings and weekends. Some have a variable pattern that shifts with school schedules, family obligations, or secondary income sources. What matters for forecasting is not just that you were active during a certain window, but how many completed deliveries per active hour you were producing during that window.
Order density varies by time of day, day of week, and neighborhood. A courier who is active in Pinheiros between 11:30am and 2pm on weekdays is operating in a high-density lunch window in one of Sao Paulo's wealthiest food-order neighborhoods. A courier who is active in a residential outer-zone neighborhood on Tuesday mornings faces a fundamentally different demand environment. Both can produce reasonable income, but the forecast model has to account for the density characteristics of each courier's actual working pattern, not just their hours.
When a courier's active windows shift, even slightly, the density signal changes. Connecting multiple weeks of history allows the model to separate stable active patterns from one-off deviations. A courier who normally earns well on Friday evenings but took one Friday off is not a different type of worker. The model has to be robust to those interruptions without over-weighting them.
Signal Two: Platform Mix and Rate Structure
Couriers who work across multiple platforms earn differently than those who work a single platform, and not always better. Each platform has its own rate structure, peak bonus architecture, and geographic concentration of orders. The mix of platforms a courier uses is therefore a meaningful input into the forecast.
A courier who is split roughly 60/40 between iFood and Rappi, for example, has a different earnings profile than one who is 90 percent iFood. The variance in per-delivery rates across the two, combined with their different peak bonus windows, means the forecast has to account for the weighted contribution of each platform based on that courier's observed usage pattern.
Platform mix also affects earnings stability. Couriers working multiple platforms tend to show lower week-to-week income variance because platform-specific slow periods get partially offset by activity on the other platform. That stability signal feeds back into the credit profile calculation as well, which is one reason multi-platform couriers in our early-access group tended to build documented income track records more quickly than single-platform couriers working the same hours.
Signal Three: Shift Timing Relative to Platform Peak Windows
The most important single behavioral predictor we found is not how much a courier works but how well their active windows align with their platforms' peak demand windows. Platforms push higher per-delivery rates and bonuses during specific demand surges, typically lunch, dinner, and weekend late-night windows. A courier who is consistently active during peak windows earns materially more per hour than one working the same total hours spread across off-peak periods.
This alignment is something that good couriers develop intuitively over months of experience, but most have no systematic data confirming whether their instincts are accurate. One of the things the Trampay forecast shows is a day-by-day projection based on expected peak window alignment given the courier's observed work pattern. The goal is not to tell couriers what to do but to make visible a relationship that is otherwise felt rather than measured.
In our early-access group, the couriers who made the most use of shift timing data reported adjusting their start times by roughly 30 to 60 minutes in one or two sessions per week. The income effect of that adjustment was meaningful at the margin, typically in the R$ 80 to R$ 200 per week range depending on their baseline hours, because they were now inside the bonus window instead of starting just after it closed.
Signal Four: Seasonal and Neighborhood-Level Variation
Income for any given courier is not just a function of their behavior. It is also a function of demand conditions in the areas where they operate. Neighborhood-level order density is not constant across the calendar year, and the patterns differ by area type.
Commercial neighborhoods like Faria Lima and Paulista see strong lunch demand during weekdays tied to office populations. Those patterns flatten during January, when Sao Paulo offices thin out, and during school holiday periods. Residential neighborhoods with dense apartment buildings see elevated weekend dinner demand throughout the year but with spikes around major events and holidays.
Seasonal variation is the hardest signal to calibrate on short data windows. A courier who joined Trampay in August and has four months of history going into December will have their first December forecast informed partly by what comparable couriers in similar work patterns experienced in prior years. We are transparent that forecasts extending into new seasonal territory carry wider uncertainty ranges than forecasts for periods well within the courier's documented history.
What the Model Does Not Know
Forecast models are bounded by their inputs. Trampay reads historical delivery data from connected platforms. We do not know about income from platforms the courier has not connected. We do not know about cash income from separate work. We do not know about planned changes in work patterns, illness, motorbike repairs, or any other future event that will affect actual hours worked.
The forecast is a probability-weighted estimate based on observed past behavior in documented conditions. It is not a guarantee of future earnings. We show confidence intervals on the forecast display rather than single-point estimates because a range that reflects actual model uncertainty is more honest than a single number that implies false precision.
The forecast is most accurate for couriers with at least three months of connected history operating in relatively stable work patterns. For couriers with less history or highly irregular patterns, the forecast window is narrower and the confidence ranges are wider. That is the honest version of what the model can and cannot do.
What it can do well is turn a subjective sense of "I think next week will be around R$ 900" into a structured estimate grounded in the actual behavioral and platform signals that have predicted that courier's income in the past. That structure is useful even when the forecast is imperfect, because planning from an informed estimate beats planning from no information at all.
