signalflow-trading 0.8.3

SignalFlow: modular framework for trading signal generation, validation and execution

SignalFlow

Real-time-first framework for trading signal research and execution

---

SignalFlow is a Polars-backed framework for algorithmic trading research that takes a strategy from idea to backtest to live with one object you can save and ship. The public surface is six nouns:

Noun	What it is
Dataset	One lazy, immutable market-data container. sf.data(...) builds it; the same object feeds backtest, paper, and live.
Transform	A column-producing step - features (SMA in core; RSI, ATR, ZScore, … via the signalflow-ta plugin) and detectors (SmaCrossDetector, ThresholdDetector) share one contract.
Models	ForecastModel (trainable predictor → probability column) plus validator combinators.
Flow	The central, deployable, tradeable unit: forecasts → detectors → validator → strategy → risk.
Engine	The decision/execution loop and brokers (SimBroker for backtest/paper, BinanceBroker for armed live).
Run	The result of executing a Flow - equity curve, fills, and a standard .scorecard().

Install

pip install signalflow-trading

Extra	Installs	For
signalflow-trading[ta]	signalflow-ta	256 technical-indicator features + 30 detectors
signalflow-trading[labs]	signalflow-labs[rl]	neural encoders, RL strategy, torch backends
signalflow-trading[live]	mlflow, huggingface_hub	model artifact tracking / deploy
signalflow-trading[llm]	httpx, pydantic	LLM-assisted strategy (any OpenAI-compatible server)
signalflow-trading[all]	ta + labs + live + llm	everything
signalflow-trading[dev]	pytest, ruff, mypy	development

Idea → first backtest

import signalflow as sf

ds = sf.data("memory", pairs=["BTCUSDT"], start="2023-01-01", interval="1h")

model = sf.ForecastModel(target=sf.FixedHorizon(bars=12),
                         features=sf.FeaturePipe(sf.SMA(10), sf.SMA(20), sf.SMA(50)))
model.fit(ds)                                          # train tier-1 forecaster

flow = sf.Flow(name="sma_rise",
               forecasts={"rise": model},
               detectors=[sf.ThresholdDetector(forecast="rise", p_min=0.6)],
               strategy=sf.RulesStrategy())
run = flow.backtest(ds, capital=50_000)
print(run.scorecard())                                 # total_return, sharpe, max_drawdown, ...

Backtest → paper → live

One decision core drives all three modes. Backtest and paper replay a finished Dataset; live consumes a streaming feed (PollingFeed polls Binance for each closed bar) and routes orders to a real venue when armed=True.

flow.paper(ds, capital=50_000)                         # sim fills over a Dataset

feed = sf.PollingFeed(sf.BinanceSource(), pairs=["BTCUSDT"], interval="1m")
flow.live(feed, capital=50_000)                         # live data, SimBroker (paper)
flow.live(feed, capital=50_000, armed=True,             # real orders on Binance
          broker=sf.BinanceBroker(api_key=..., api_secret=...),
          state_path="book.json")                       # book persists across restarts

CLI

sf list                 # registry snapshot grouped by type
sf list transform       # one type, with one-line summaries
sf run flow.yaml --source memory --pairs BTCUSDT --start 2023-01-01 --interval 1h --capital 50000
sf promote flow.yaml --to shadow   # validate + show the registry op (real promotion: sf-prod)
sf version

Three invariants worth knowing

WoE/IV encoding is the default. Features flow through Weight-of-Evidence encoding against the target, and IVSelector keeps only columns whose Information Value clears a threshold. Encoding is monotone, leak-aware, and fit out-of-fold.

A Flow is inference-only. Every forecast slot (and the optional validator slot) must hold a trained model. Constructing a Flow around an unfitted model raises UntrainedModelError - you cannot accidentally deploy something untrained. The same Flow object runs backtest and paper over a Dataset and live over a streaming feed - one decision core, no separate execution path to drift.

Deploy is data. flow.save(path) serializes the whole stack (config + trained artifacts) to YAML plus a model directory; sf.Flow.load(path) brings it back byte-for-byte. There is no code to redeploy - promoting a strategy is moving a file. Model artifacts can live on the local filesystem, MLflow, or the Hugging Face Hub (model.save("mlflow://..."), model.save("hf://...")).

flow.save("flows/rsi_rise.yaml")
same = sf.Flow.load("flows/rsi_rise.yaml")
assert same.backtest(ds, capital=50_000).final_equity == run.final_equity

Registry

Every core class registers under a name - that name is what flow.yaml serializes and sf list enumerates. Seven ComponentTypes: SOURCE, TRANSFORM, MODEL, STRATEGY, SAMPLER, BROKER, METRIC.

sf.registry.snapshot()                          # {type: [names]}
sf.registry.list(sf.ComponentType.TRANSFORM)    # core: ['sma', 'woe', ...]; +256 with [ta]

Installing a plugin (signalflow-ta, signalflow-labs) auto-registers its components via entry points - no imports or wiring needed.

Ecosystem

Package	Description
signalflow-ta	Technical-indicator plugin: 256 features + 30 detectors ([ta] extra)
signalflow-labs	Neural encoders, RL strategy, torch backends ([labs] extra)
sf-prod	Promotion, shadow/live rollout, monitoring

---

License: MIT  ·  Author: pathway2nothing  ·  Docs: signalflow-trading.com