Fast-and-frugal trees
A short decision tree with an exit at every question, for quick high-stakes calls.
What it means
A fast-and-frugal tree is a simple classification heuristic: a sequence of binary cues in which each cue offers an immediate exit to a decision, so a verdict can be reached after as few as one question. Unlike a full decision tree it has the minimum number of branches — one exit per node — making it transparent, memorable, and fast under pressure. Despite their austerity, such trees can rival statistical classifiers in predictive accuracy while being far easier for practitioners to apply and trust, again reflecting robustness under limited data. The ordering and exit structure of cues can be tuned to balance misses against false alarms for a given context. They are used in emergency medicine, screening, and triage where speed, interpretability, and cost all matter.
How they are built
A fast-and-frugal tree is not guessed; it is constructed from data. Each cue needs a threshold — for a numeric cue, often its median — and an exit direction. Simple methods rank cues by marginal predictive validity: the 'max' rule orders them by highest validity, 'zig-zag' alternates the exit direction down the tree. Newer 'fan' algorithms grow several candidate trees with different exit structures and keep the one that best matches the decision-maker's desired error trade-off; one variant re-estimates thresholds on the cases each branch actually receives, letting it catch some cue interactions. In one benchmark these fan-built trees used fewer than two cues on average yet beat the older rules on accuracy. The output stays legible: a clinician or clerk can read the whole rule off a single page.
What the evidence shows
The founding case is Green and Mehr's Michigan coronary-care study. A three-question tree — an ECG anomaly, then chest pain as the chief complaint, then a small set of secondary factors — matched a regression-based instrument that weighed seven cues, and sent fewer low-risk patients to intensive care than the physicians' unaided judgment did. Broader tests hold up. Across thirty classification problems Martignon and colleagues found simple trees close to CART and logistic regression; across ten datasets Phillips and colleagues report a mean balanced accuracy near 0.83, matching regularized logistic regression and random forests and trailing support-vector machines only slightly. Their edge is robustness: with few cues there is little to overfit, so training and test accuracy stay close where more elaborate models diverge.
Tuning misses against false alarms
Luan, Schooler, and Gigerenzer showed the design choices map cleanly onto signal-detection theory. The exit structure sets the tree's response bias: a tree whose early exits all point to 'act' rarely misses a true case but raises false alarms, while early 'clear' exits do the reverse. Where a miss is far costlier than a false alarm — a missed heart attack, an undetected fraud — you rake the exits toward action; where false alarms are expensive you rake the other way. Cue order and cue quality, by contrast, set the tree's sensitivity, its raw power to tell cases apart. This separation is what lets a practitioner deliberately choose an operating point rather than accept whatever a black-box classifier happens to return.
Where they break down
The frugality has costs. Standard algorithms treat cues as independent, so a tree can miss patterns that live in the interaction between cues. They fare worst exactly where linear models shine: many individually weak cues whose weighted sum is strongly predictive, information a single-exit tree throws away. Construction usually ignores what each cue costs to obtain, so a tree may put an expensive test before a cheap one of similar value unless the order is fixed by hand. And the very act of ignoring available data can be hard to defend after a bad outcome, even when the simpler rule is more accurate on average — a governance problem more than a statistical one. Frugality also cannot rescue a poor set of cues.
Examples
An emergency physician decides whether to send a chest-pain patient to coronary care by checking a few ordered signs, sending the patient at the first cue that is positive.
A loan officer without a scorecard asks three ordered questions, missed payments, verified income, debt below a limit, and can decline at the first answer, approving only what clears them all.
A recruiter screening hundreds of CVs rejects at the first no and puts the cheapest question first; the order is tuned to trade a few good candidates missed against hours saved.
A wildfire crew boss decides on evacuation by asking, in order: is the fire within range, is the wind toward town, is the road open, acting the moment one answer forces the call.
A line inspector routes a part to scrap at the first failed check, a visible crack, and passes it to shipping only if it clears the crack, weight, and seal tests in turn.
First described in Green & Mehr (1997); Gigerenzer & colleagues.
Key references
- Martignon, L., Erickson, T., & Viale, R. (2022). Transparent, simple and robust fast-and-frugal trees and their construction. Frontiers in Human Dynamics, 4, 790033. doi.org/10.3389/fhumd.2022.790033
- Phillips, N. D., Neth, H., Woike, J. K., & Gaissmaier, W. (2017). FFTrees: A toolbox to create, visualize, and evaluate fast-and-frugal decision trees. Judgment and Decision Making, 12(4), 344-368. doi.org/10.1017/S1930297500006239
- Luan, S., Schooler, L. J., & Gigerenzer, G. (2011). A signal-detection analysis of fast-and-frugal trees. Psychological Review, 118(2), 316-338. doi.org/10.1037/a0022684
- Martignon, L., Katsikopoulos, K. V., & Woike, J. K. (2008). Categorization with limited resources: A family of simple heuristics. Journal of Mathematical Psychology, 52(6), 352-361. doi.org/10.1016/j.jmp.2008.04.003
- Green, L., & Mehr, D. R. (1997). What alters physicians' decisions to admit to the coronary care unit? Journal of Family Practice, 45(3), 219-226. pubmed.ncbi.nlm.nih.gov/9300001/