A decision tree divides data by asking a sequence of questions. A shallow tree can only ask a few questions along a path. A deeper one can make increasingly narrow distinctions, including distinctions that describe quirks of the training sample.
The scikit-learn decision-tree regression example illustrates how greater depth can fit noise as well as the underlying pattern. More detail in a training prediction is therefore not enough to establish a better model.
Use depth as a controlled comparison
Start with one prepared dataset and a split that suits the problem. Time-dependent records may require a chronological split; repeated records from the same entity may need to stay together. Otherwise, a flattering validation score can hide information leakage.
Fit a small set of candidate depths while keeping the other settings and evaluation procedure consistent. Record training and validation performance side by side. If training performance keeps improving while validation performance worsens, the extra branches may be learning detail that does not travel well.
This is an experiment to run on your data, not a promise that one particular depth will win. Dataset size, noise, feature quality and the task all affect the result.
Inspect what the branches are doing
Read a few paths from root to leaf. Ask whether the split variables would be available at prediction time and whether tiny groups are producing surprisingly extreme outputs. A technically valid tree can still reveal a flawed feature or a data-preparation mistake.
Depth is only one control. Minimum samples per leaf and other constraints also affect how the tree divides the data. Change them deliberately rather than treating a large parameter search as a substitute for understanding the model.
Keep the final test data separate from model selection. Report the evaluation method alongside the chosen settings, and retain the simpler candidate if added complexity produces no dependable benefit. The useful tree is the one that answers the task reliably, not the one with the most branches to display.