Release 0.11.0 — Machine Learning, Kolmogorov-Arnold Networks (KAN), and Technical Indicator Networks (TIN)
Date: 2026-11-16
Install: pip install quantwave or pip install "quantwave[polars]"
Headline
QuantWave 0.11.0 introduces a production-grade Machine Learning & Neural Indicators subsystem directly into its high-performance Rust core (quantwave-core::ml).
Departing from opaque, over-parameterized deep learning architectures that degrade on low signal-to-noise financial data, 0.11.0 delivers two research-backed paradigms specifically designed for quantitative trading: 1. Kolmogorov-Arnold Networks (KAN): Non-linear function synthesis via learnable univariate B-splines on network edges with closed-form symbolic mathematical rule extraction (KASPER, TMLR 2026). 2. Technical Indicator Networks (TIN): Differentiable, learnable technical indicators with exact recurrence analytical parameter derivatives (Lu, arXiv:2507.20202).
All implementations are 100% safe Rust (unsafe_code = "forbid"), zero-heap-allocation on the streaming hot path, and fully integrated with QuantWave's universal streaming traits (Next<T> and SmoothingAlgorithm).
What's New
1. Kolmogorov-Arnold Networks (KAN)
BSplineBasis: Pure safe-Rust Cox-de Boor B-spline basis evaluator supporting arbitrary polynomial degrees \(k\) and uniform knot grids. Guaranteed partition of unity (\(\sum B_i = 1\)) and boundary clamping.KanLayer: Edge-parameterized neural layer mapping \(\mathbb{R}^n \to \mathbb{R}^m\) via learnable B-splines and residual \(\text{SiLU}\) activations.KanNetwork: Deep sequential KAN with pre-allocated ping-pong intermediate buffers for allocation-free multi-layer forward passes.KanIndicator: Real-time streaming wrapper implementingNext<&[f64]>andNext<f64>for live tick execution.- Symbolic Rule Extraction (
extract_symbolic_formula): Evaluates \(L_1\) coefficient norms across edges to prune inactive features and output human-readable, auditable mathematical expressions.
2. Technical Indicator Networks (TIN)
DifferentiableEma: Continuous lookback span \(\tau \in \mathbb{R}^+\) with exact analytical recurrence gradients: $\(\frac{\partial y_t}{\partial \alpha} = (x_t - y_{t-1}) + (1 - \alpha) \frac{\partial y_{t-1}}{\partial \alpha}\)$ Supports online stochastic gradient descent viaupdate_spanand implementsSmoothingAlgorithm.DifferentiableRsi: Dual-EMA tracking for gain and loss with analytical span derivatives and smooth sigmoidal decision boundaries in \([-1.0, 1.0]\).DifferentiableMacd: Multi-span continuous MACD with analytical gradients for fast, slow, and signal spans.
3. Python API & PyTorch Integration (quantwave.ml)
KanNetwork&KanIndicator: Python bindings with.forward(),.forward_batch(), streaming.next()/.update(), and JSON serialization (to_json/from_json).- Symbolic Distillation: Python functions
extract_symbolic_formula()andextract_feature_importance(). - Differentiable Indicators:
DifferentiableEma,DifferentiableRsi, andDifferentiableMacdwith.step(),.update_span(), and.batch()methods. - PyTorch Autograd Bridge: Native
LearnableEmaandLearnableRsinn.Moduleclasses using Rust analytical recurrence gradients for backpropagation in PyTorch.
4. Documentation & Guides
- Added comprehensive guide in
docs/guides/machine_learning.mddetailing individual mathematical foundations, API structures, streaming Rust examples, Python (quantwave.ml) usage guidelines, and the two-stage "TIN-KAN" hybrid pipeline.