Skip to content

Quick Start

The problem

Most quant libraries compute indicators as a batch over a full DataFrame. That means your research code and your live trading code are fundamentally different objects. Any inconsistency between them is a live trading bug waiting to happen.

The Oryon model

Every feature is a stateful object. It keeps just enough history to compute its output. You feed it one bar at a time, or the whole dataset. Same object, same logic, same output.

from oryon import FeaturePipeline
from oryon.features import Sma

sma = Sma(["close"], window=20, outputs=["close_sma_20"])
fp  = FeaturePipeline(features=[sma], input_columns=["close"])

# Live: one bar at a time
result = fp.update([109.5])   # [nan] during warm-up, [value] once ready

# Research: same object, same code
results = fp.run_research([[100.0], [101.5], [103.0], ...])

There is no separate research implementation. The same Rust code runs in both contexts.

Build a dataset in 10 lines

import pandas as pd
from oryon import FeaturePipeline, TargetPipeline
from oryon import run_features_pipeline, run_targets_pipeline, load_sample_bars
from oryon.features import Ema, LogReturn, ParkinsonVolatility
from oryon.targets import FutureReturn

df = load_sample_bars()  # ~14 000 OHLCV bars, ready to use

fp = FeaturePipeline(
    features=[
        Ema(["close"], window=20, outputs=["ema_20"]),
        LogReturn(["close"], window=1, outputs=["log_ret"]),
        ParkinsonVolatility(["high", "low"], window=20, outputs=["pvol_20"]),
    ],
    input_columns=["close", "high", "low"],
)
tp = TargetPipeline(
    targets=[FutureReturn(inputs=["close"], horizon=5, outputs=["ret_5"])],
    input_columns=["close"],
)

X = run_features_pipeline(fp, df)
y = run_targets_pipeline(tp, df)
dataset = pd.concat([X, y], axis=1).dropna()

The same fp pipeline can then be used in live trading. Call fp.update(bar) on every new tick. No rewrite, no translation layer.

Performance

Features run under 1µs per update in Python at window=200 (measured via pytest-benchmark on Apple M-series):

Feature Median (w=200)
EMA, SimpleReturn, LogReturn <200ns
Parkinson, Rogers-Satchell, SMA <500ns
LinearSlope, Skewness, Kurtosis <1µs
KAMA ≈1µs

See the full table on the Benchmarks page.

Next