If your team is forecasting merchant volume, predicting chargeback risk, or building any kind of model to guide portfolio decisions, there’s one diagnostic that matters more than almost any summary statistic you’ll pull: the residual. It’s simple, it’s often skipped, and it’s usually the fastest way to tell whether a model can be trusted with real decisions.
Here’s what residuals are, why they matter for the kind of forecasting partners rely on every day, and what to look for when you check them.
What a residual actually is
A residual is the leftover. For any single data point, it’s the gap between what actually happened and what your model predicted would happen.
Residual = Actual value − Predicted value
Say you’re forecasting a merchant’s monthly processing volume based on historical trends and transaction counts. The model predicts $180,000 for the month. Actual volume comes in at $192,000. The residual is $12,000. Your model undershot by that amount.
Run this across every merchant or every period in your dataset and you get a full set of residuals. That’s where the real diagnostic work starts, and it’s the part most people skip past on their way to a summary number.
Why residuals matter more than a single accuracy score
It’s tempting to look at one number, like overall forecast accuracy, and call the model good enough. The problem is that a strong average can hide bad behavior underneath it.
A model can look accurate overall and still be systematically wrong in the exact places that matter most to your portfolio, like underestimating volume for your highest-risk merchant segment or overestimating it for a category that’s actually shrinking. Residuals surface that, because you’re looking at the full pattern of errors instead of just their average size.
What to actually look for
The most useful move is to plot the residuals. Predicted values on one axis, residuals on the other, then look at the shape.
A healthy residual plot looks like random noise. Points scattered evenly above and below zero, no visible pattern, no trend. That randomness is the goal. It means the model has already pulled the real signal out of the data, and what’s left is genuinely unpredictable, not something the model missed.
An unhealthy residual plot tells you something specific:
- A curve in the residuals usually means the model is missing a nonlinear relationship, like volume that accelerates seasonally instead of growing at a steady rate.
- A funnel shape, where residuals widen as predicted values increase, means your model’s errors get less reliable for larger merchants or higher volumes. That’s a real problem if you’re using the same model to size risk across small and large accounts alike.
- Clusters or bands of residuals often point to a missing variable, like merchant category, region, or processor type, that the model isn’t accounting for.
- A sloped trend in the residuals means the model is systematically biased in one direction. That’s the kind of bias that quietly skews portfolio decisions if nobody catches it.
A quick way to think about it
Residuals are the model’s exit interview. Every forecast gets checked against what actually happened, and the residual is that feedback. Random, scattered feedback means the model is doing its job. Patterned feedback means the model is missing something, and the pattern itself is usually a clue about what that something is.
Where to check this
Nothing exotic is required here:
- Residuals vs. predicted values: the core check described above
- Residuals vs. a specific input, like merchant category or region: sometimes the overall plot looks clean, but a pattern only shows up when you isolate one variable
- Histogram of residuals: a fast way to see if errors skew consistently in one direction
So before that forecast drives a portfolio call or a risk decision, pull up the residuals. If they scatter randomly with no shape to them, that’s a good sign, take the win. If there’s a pattern, don’t file it away. That pattern is telling you exactly what the model missed and where to go fix it.