# Oryon

Rust library for feature and target engineering in quantitative trading.
Python API via PyO3/maturin. Version 0.2.x.

## What it does

Oryon builds ML datasets from financial time series (OHLCV bars).
Three types of streaming transforms (all implement StreamingTransform, run live + research):
- Features: stateful indicators with internal rolling state.
- Scalers: normalization transforms, rolling or pre-fitted.
- Operators: stateless math combining or transforming feature outputs.

Plus batch labels for research:
- Targets: forward-looking labels. Require future bars, research-only.

Every StreamingTransform runs identically in live trading and backtesting.
No separate research implementation.

## Features (v0.2)

- Trend: Sma, Ema, Kama
- Returns: SimpleReturn, LogReturn
- Statistics: Adf, Skewness, Kurtosis, LinearSlope, Mma, ShannonEntropy, Correlation, AutoCorrelation
- Volatility: ParkinsonVolatility, RogersSatchellVolatility

## Scalers (v0.2)

- RollingZScore: rolling (x - mean) / std over a sliding window. Live-compatible.
- FixedZScore: (x - mean) / std with pre-fitted parameters. Use with fit_standard_scaler.

## Operators (v0.2)

- Add: A + B. Returns None if either input is None.
- Subtract: A - B. Returns None if either input is None.
- Multiply: A * B. Returns None if either input is None.
- Divide: A / B. Returns None if either input is None or B is 0.
- Reciprocal: 1/x. Returns None if x is None or 0.
- Log: ln(x). Returns None if x <= 0 or None.
- NegLog: -ln(x). Returns None if x <= 0 or None.
- Logit: ln(x/(1-x)). Returns None if x outside (0, 1) or None.

## Fitting (v0.2)

- fit_standard_scaler(data): computes mean and std from a column. Returns (mean, std).
  Skips None values. Use the result to construct FixedZScore.

## Targets (v0.2)

- FutureReturn
- FutureCTCVolatility
- FutureLinearSlope

## Pipelines

- FeaturePipeline: runs a DAG of StreamingTransforms (features, scalers, operators) on a
  dataset or one bar at a time. Dependencies resolved automatically by output/input names.
- TargetPipeline: runs a list of targets on a full dataset (research only).

## Key conventions

- Missing values are Option<f64> in Rust, NaN in Python. Never sentinel floats.
- warm_up_period: number of NaN bars at the start of a StreamingTransform output.
- forward_period: number of NaN bars at the end of a target output.
- reset() clears internal state. Call between walk-forward folds.
- Operators have warm_up_period = 0 and no internal state.
- FixedZScore has warm_up_period = 0 (params are pre-fitted).

## Datasets

- load_sample_bars(): ~14 000 OHLCV bars (time, open, high, low, close, volume), embedded in the package.

## Install

pip install oryon

## Links

- Docs: https://oryonlib.dev
- Source: https://github.com/lucasinglese/oryon
- PyPI: https://pypi.org/project/oryon