signalflow-labs

Neural-network & RL extension for SignalFlow - 14 encoders, 7 heads, 4 losses, RL strategy

Part of the SignalFlow ecosystem.

A plugin: a PyTorch/Lightning library for financial time-series classification (modular encoders, classification heads, loss functions) plus a reinforcement-learning strategy. Installing it auto-registers its components with the core signalflow registry via entry points.

Installation

pip install signalflow-labs           # core: torch + lightning
pip install "signalflow-labs[rl]"     # + stable-baselines3, gymnasium (RL strategy)
# or, from the core:
pip install "signalflow-trading[labs]"

Requires: Python ≥ 3.12, signalflow-trading ≥ 0.8.0, PyTorch ≥ 2.2, Lightning ≥ 2.5.

Quick Start

from signalflow.labs.encoder import TransformerEncoder
from signalflow.labs.head import MLPClassifierHead
from signalflow.labs.model import TemporalClassificator
from signalflow.labs.data import SignalDataModule
import lightning as pl

# Create model
model = TemporalClassificator(
    encoder_type="encoder/transformer",
    encoder_params={"d_model": 64, "nhead": 4, "num_layers": 2},
    head_type="head/cls/mlp",
    head_params={"hidden_sizes": [32]},
    num_classes=3,  # fall, neutral, rise
)

# Create data module
dm = SignalDataModule(
    data=df,
    window_size=60,
    batch_size=32,
    split_strategy="temporal",
)

# Train
trainer = pl.Trainer(max_epochs=50, accelerator="auto")
trainer.fit(model, dm)

Encoders (14)

Classification Heads (7)

Loss Functions (4)

SignalFlow Integration

Importing signalflow.labs registers its components. The RL strategy plugs into a Flow as the strategy slot (needs the [rl] extra):

import signalflow as sf
import signalflow.labs as labs               # registers neural + RL components
from stable_baselines3 import PPO

base = sf.Flow(name="rl", detectors=[sf.SmaCrossDetector()])
env = labs.make_env(base, ds)                # gymnasium env over an Engine replay
policy = PPO("MlpPolicy", env).learn(10_000)

flow = base.replace(strategy=labs.RLStrategy(model=policy, size_pct=0.1))
run = flow.backtest(ds, capital=50_000)
print(run.scorecard())

The neural TemporalClassificator can also back a ForecastModel for signal validation; see signalflow.labs.validator.

Package Structure

License: MIT  ·  Part of SignalFlow