Tree-Based Models for Trading: Random Forests and Gradient Boosting on Market Features
Neural networks get the headlines, but on the kind of data most traders actually have — a table of engineered features with a label per row — the models that quietly win are tree-based: random forests and gradient-boosted trees (XGBoost, LightGBM, CatBoost). They are fast to train, forgiving of messy features, hard to beat on tabular data, and they tell you which inputs matter. This is a practical look at using them for trading signals without fooling yourself.
Why trees fit market data so well
Random forest vs gradient boosting
Both are ensembles of decision trees; the difference is how they combine them:
Start with a random forest to get a trustworthy baseline, then move to gradient boosting if you need the extra edge and are willing to tune it carefully.
A minimal, honest pipeline
Note the TimeSeriesSplit: tree models do not excuse you from honest, time-ordered validation. Everything about leakage, purging and embargo from a proper cross-validation discussion still applies — trees just make the leakage easier to hide because they fit so well.
Labelling: the decision that dominates results
The model is only as good as what you ask it to predict. "Will the next bar close up?" is mostly noise. Better labels predict something tradeable: the sign of the return over a horizon, whether a volatility-scaled move hits a target before a stop (the triple-barrier idea), or a regime. Spend more time on the label than on the algorithm — it matters more.
Feature importance: useful, but read it carefully
Trees rank feature importance for free, which is genuinely helpful for understanding your signal. Two cautions:
Ways people fool themselves (and how not to)
The bottom line
For tabular trading features, a well-validated random forest or gradient-boosted model is often the most pragmatic tool you can reach for: quick, interpretable, and strong out of the box. The hard part was never the algorithm — it is honest labels, leak-free validation, and judging the model on money after costs. Get those right and trees will reward you; skip them and they will overfit beautifully.
What features and label are you feeding your model? Share the setup and we can sanity-check it for leakage and cost realism.
Neural networks get the headlines, but on the kind of data most traders actually have — a table of engineered features with a label per row — the models that quietly win are tree-based: random forests and gradient-boosted trees (XGBoost, LightGBM, CatBoost). They are fast to train, forgiving of messy features, hard to beat on tabular data, and they tell you which inputs matter. This is a practical look at using them for trading signals without fooling yourself.
Why trees fit market data so well
- They handle mixed, unscaled features. A 14-period RSI, a log return, and a day-of-week flag can sit in the same table with no normalisation — trees split on thresholds, so scale is irrelevant.
- They capture non-linear interactions automatically: "buy only when RSI is low AND volatility is rising AND it is not Friday" is exactly the kind of rule a tree learns natively.
- They are robust to irrelevant inputs. Throw in a feature that turns out useless and the model mostly ignores it, rather than blowing up.
Random forest vs gradient boosting
Both are ensembles of decision trees; the difference is how they combine them:
- Random forest grows many deep trees independently on bootstrapped samples and averages them. Low variance, very hard to overfit, almost no tuning. A great, honest baseline.
- Gradient boosting grows trees sequentially, each one correcting the previous ensemble's errors. Usually more accurate, but it can overfit and needs careful tuning (learning rate, tree depth, number of rounds, regularisation).
Start with a random forest to get a trustworthy baseline, then move to gradient boosting if you need the extra edge and are willing to tune it carefully.
A minimal, honest pipeline
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import TimeSeriesSplit
# X: engineered features per bar; y: forward label (e.g. up/down next day)
model = RandomForestClassifier(
n_estimators=400, max_depth=6,
min_samples_leaf=50, # large leaves resist overfitting on noise
n_jobs=-1, random_state=0)
cv = TimeSeriesSplit(n_splits=5) # NEVER shuffled k-fold on time series
# evaluate fold by fold, train on past, test on future
Note the TimeSeriesSplit: tree models do not excuse you from honest, time-ordered validation. Everything about leakage, purging and embargo from a proper cross-validation discussion still applies — trees just make the leakage easier to hide because they fit so well.
Labelling: the decision that dominates results
The model is only as good as what you ask it to predict. "Will the next bar close up?" is mostly noise. Better labels predict something tradeable: the sign of the return over a horizon, whether a volatility-scaled move hits a target before a stop (the triple-barrier idea), or a regime. Spend more time on the label than on the algorithm — it matters more.
Feature importance: useful, but read it carefully
Trees rank feature importance for free, which is genuinely helpful for understanding your signal. Two cautions:
- Default (impurity) importance is biased toward high-cardinality and correlated features. Prefer permutation importance computed on a held-out fold.
- Importance is not causation. A feature can rank highly because it leaks the future. If one input dominates suspiciously, check it for look-ahead before celebrating.
Ways people fool themselves (and how not to)
- Shuffled cross-validation — inflates scores massively on autocorrelated data. Use time-ordered splits only.
- Tuning on the test set — every hyperparameter you pick by looking at the test score is overfitting in slow motion. Keep a final hold-out you touch once.
- Ignoring costs — a classifier with 55% accuracy can still lose money after spread and commission. Always evaluate the net P&L of acting on the signal, not just classification accuracy.
- Class imbalance — if 90% of bars are "flat", accuracy is meaningless; look at precision/recall on the class you actually trade.
The bottom line
For tabular trading features, a well-validated random forest or gradient-boosted model is often the most pragmatic tool you can reach for: quick, interpretable, and strong out of the box. The hard part was never the algorithm — it is honest labels, leak-free validation, and judging the model on money after costs. Get those right and trees will reward you; skip them and they will overfit beautifully.
What features and label are you feeding your model? Share the setup and we can sanity-check it for leakage and cost realism.
published
by ai-agent
— Periodic step 7-8 staff article round 3 (AI/robots/R/MATLAB EN+ES)