AmeforgeAmeforge
PublishedOpen Source · CKLarXiv:2603.14651

EARCP

Ensemble Auto-Regulative Coherence Protocol

EARCP is a novel ensemble learning framework that dynamically combines multiple expert models by weighting them according to both their recent predictive performance and their inter-model coherence. Unlike simple averaging, EARCP adapts continuously to non-stationary environments where data distributions shift over time.

arXiv:2603.14651
March 15, 2026
cs.LG · cs.AI · cs.MA
Ameforge Research Lab
Independent AI Research
contact@ameforge.tech
github.com/Volgat/earcp
Language: Python / PyTorch
License: CKL · PyPI available

Ensemble models fail in non-stationary environments

Traditional ensemble methods assign static weights to their constituent models. When the underlying data distribution shifts, as it does in financial markets, industrial sensors, or real-time systems. These fixed weights become obsolete, degrading performance without any mechanism for self-correction.

Simple averaging treats all models equally regardless of their recent accuracy, while manual reweighting requires continuous human intervention and domain expertise.

Dynamic weighting via performance and coherence

EARCP adjusts the weight of each expert model at every prediction step based on two signals: recent predictive performance (who has been right lately) and inter-model coherence (who agrees with the consensus).

Key insight: Coherence prevents the algorithm from over-relying on a model that was correct by chance but is isolated from the consensus of other experts. This provides a theoretical guarantee of bounded regret over time.

Three expert models, one adaptive coordinator

C

CNN Expert

Pattern Recognition

Convolutional layers extract local temporal patterns and short-range dependencies from the input sequence.

B

BiLSTM Expert

Sequence Modeling

Bidirectional LSTM captures long-range temporal dependencies in both forward and backward directions.

T

Transformer Expert

Attention Mechanism

Self-attention mechanism models global relationships across the entire input sequence simultaneously.

# Weight update rule (from source code, line 138)
score = beta * (-performance) + (1 - beta) * coherence
weights = softmax(eta * score)
# Where:
beta = balance between performance and coherence
eta = learning rate for weight updates
softmax = normalized weight distribution across experts

Hull Tactical Market Prediction Competition

708
Final Rank
Public leaderboard
18 000
Participants
International competition
1.12
Sharpe Ratio
Risk-adjusted return
+43%
vs. Simple Average
Sharpe improvement

Configuration: EARCP with 3 experts (CNN, BiLSTM, Transformer). Max drawdown: -8.3%. Evaluation period: 8 months on live market data

Verifiable: Results are publicly visible on the Kaggle leaderboard under the profile volgat. The competition is organized by Hull Tactical Asset Allocation.

kaggle.com · Hull Tactical Leaderboard
Kaggle Hull Tactical Leaderboard, Rank 708 (volgat)

Kaggle leaderboard search result for "Volgat": Rank 708, Score 0.888

The Hull Tactical Market Prediction competition evaluates algorithms on their ability to predict market predictability, a second-order forecasting task. All results reported here are from the public leaderboard and are independently verifiable. The Sharpe ratio and drawdown figures are computed by the competition platform, not self-reported.

Designed for non-stationary environments

EARCP is a general-purpose ensemble architecture applicable to any multi-model system where the data distribution changes over time.

Finance
Market regime shifts
Validated
Industrial Sensors
Sensor drift & calibration
POC in progress
Automotive
Weather & road conditions
Planned
Cybersecurity
Novel attack patterns
Under study

Get started with EARCP

1. Install via PyPI

bash
pip install earcp

2. Basic usage

python
from earcp import EARCP

# Initialize with your expert models
model = EARCP(experts=[cnn, lstm, transformer])

# Make a prediction
prediction = model.predict(x)

# The weights are updated automatically
# based on performance and coherence

3. Advanced configuration

python
from earcp import EARCP

model = EARCP(
    experts=[cnn, lstm, transformer],
    beta=0.5,      # balance: performance vs coherence
    eta=0.1,       # learning rate for weight updates
    window=20      # lookback window for performance
)

# Access current weights
print(model.weights)  # [0.42, 0.31, 0.27]

# Evaluate on benchmark data
results = model.evaluate(X_test, y_test)
Requirements
Python >= 3.8
PyTorch >= 1.12
NumPy >= 1.21
scikit-learn >= 1.0