The example script (notebooks/example_triple_barrier_walkthrough.py) exposes the model search near the top, so you can change the experiment without editing the training loop:
LGBM_SEARCH_GRID = {
'n_estimators': [25, 50, 75, 100],
'num_leaves': [7, 15, 31],
'max_depth': [3,4,5],
'learning_rate': [0.01, 0.03, 0.05],
}
# Shared settings; a parameter in the search grid overrides its fixed value.
LGBM_FIXED_PARAMS = dict(
# learning_rate=.03,
# max_depth=5,
min_child_samples=30,
random_state=42,
n_jobs=2,
verbosity=-1
)
Scikit-learn's ParameterGrid generates every combination: 4 tree counts × 3 leaf counts × 3 depths × 3 learning rates = 108 candidates. LGBM_FIXED_PARAMS applies to every candidate; a setting in the search grid overrides the corresponding fixed value.
Each candidate is evaluated on the first three folds, and the lowest mean multiclass log loss wins. Log loss evaluates the probabilities themselves: assigning high confidence to the wrong class is more costly than admitting uncertainty.
For the saved Gold run, the selected configuration was:
| Trees |
Leaves |
Maximum depth |
Learning rate |
Mean selection log loss |
| 25 |
15 |
3 |
0.03 |
1.0013 |
That configuration stays fixed for both OOS folds. The model is refitted before each fold using only labels available at its training cutoff; resolved outcomes from the earlier OOS fold can enter the later fold's training data, but never the parameter search. All performance and trading figures below combine these two folds. After evaluation, the script refits on all eligible data to create the production artifact without overwriting the OOS results.
Change the features in the same place
Immediately below the model parameters, the example exposes both a recipe list and the function that constructs features:
ENGINEERED_SPECS = [
# {'kind': 'log_return', 'window_bars': 1},
# {'kind': 'log_return', 'window_bars': 6},
# {'kind': 'log_return', 'window_bars': 24},
]
def engineer_features(frame, specs, input_bars):
"""Edit here: return (dataframe, added column names) for training AND serving.
Use only feature_* inputs, which are normalized OHLCV ratios. Compute each
row independently from its input window: serving supplies a single row.
Never use target_*/tb_* columns or shift/roll across training rows. Return
every added model feature's name. Keep imports inside this function so its
code can travel with predict.pkl; do not close over credentials/data managers.
"""
from allora_forge_builder_kit import apply_engineered_features
frame, added = apply_engineered_features(frame, specs, input_bars)
# Example custom feature (uncomment both lines to enable):
# frame['last_bar_range'] = frame[f'feature_high_{input_bars-1}'] - frame[f'feature_low_{input_bars-1}']
# added.append('last_bar_range')
return frame, added
The saved run uses the normalized OHLCV inputs with an empty ENGINEERED_SPECS list. Uncomment a log-return recipe to add it, or edit engineer_features() to implement your own transformations. The same function is captured in predict.pkl and used during live inference. Work within each row's input window: serving supplies a single row, so rolling across dataset rows would make training and inference inconsistent.
Measure the probabilities, not just the winning class
The Builder Kit's PerformanceEvaluator compares the classifier against a causal baseline made from the previous 100 resolved targets. This baseline adapts to the recent class mix without looking at future labels.
The saved run contains 950 held-out predictions, from August 22 through September 30, 2026. Worker accuracy was 53.58%, compared with 44.21% for the baseline.
| Evaluation measure |
Estimate |
Bootstrap lower–upper bounds |
Passing rule |
| Accuracy improvement |
+9.37 percentage points |
+1.58 to +17.89 pp |
Estimate > 2 pp and lower bound > 0 |
| Quadratic weighted kappa |
0.1230 |
0.0234 to 0.2166 |
Lower bound > 0 |
| Brier skill |
0.2945 |
0.2463 to 0.3416 |
Lower bound > 0 |
| Focal skill |
0.9390 |
0.9152 to 0.9525 |
Lower bound > 0 |
| Offline coverage |
100% |
— |
Coverage > 90% in this offline report |
Accuracy improvement contributes two criteria, making six in total. The bounds use paired circular block bootstrap resampling: 10-row blocks, 1,000 replicates, and the 5th and 95th percentiles. Kappa accounts for the ordering of the classes, while Brier and focal skills evaluate the probability assignments relative to the baseline.
This run passed all six offline checks. That is a result for this held-out period, not evidence of live network participation: a deployed worker still needs to submit reliably. The precise criteria and formulas are in the classification evaluation reference.
Triple barrier as a distilled trading problem
The target also defines a simple trade. At each row, use the highest-probability class as a signal:
- Up: open a long position.
- Down: open a short position.
- Neutral: do nothing.
For either directional signal, exit at the first price barrier touched. If neither is touched, exit at the close of the final minute candle inside the evaluation window. The same price path that determines the label therefore determines the illustrative trade's outcome.
