Skip to main content

Financial Machine Learning

Orion is built on the principle that financial machine learning can make portfolio management radically more accessible.

We see financial machine learning as a tool to:

  • Simplify decision-making for everyday users;
  • Enhance performance for advanced managers;
  • Automate allocation in dynamic market conditions.

To support a diverse range of user needs and intents, Orion enables two core approaches:

  • Passive portfolio construction, powered by DeFi-native indexing and onchain metrics;
  • Active, ML-driven management, where adaptive models optimize allocation and rebalancing.

This dual system lets vaults range from hands-off, rules-based strategies to fully adaptive, performance-seeking allocations — all within a unified, permissionless framework.

Passive Management: DeFi-Native Indexing​

Orion enables a permissionless marketplace for onchain indexes — passive portfolios based on predefined rulesets, rebalanced automatically, and tokenized as ERC-20s.

We are building a system where:

  • Users define index weights based on asset classes, themes, or rulesets;
  • Orion integrates APIs from skfolio to define quantitative finance metrics (e.g., covariance, tail risk, train-test split) that inform onchain vault behavior.
  • Index logic is encoded onchain via IPFS CIDs pointing to dockerized portfolio management logic, ensuring deterministic and verifiable index construction.
Example of a ruleset
 from skfolio import RiskMeasure
 from skfolio.optimization import MeanRisk, ObjectiveFunction

 X = get_universe_returns()
 model = MeanRisk(
    risk_measure=RiskMeasure.STANDARD_DEVIATION,
    objective_function=ObjectiveFunction.MAXIMIZE_RATIO,
    portfolio_params=dict(name="Max Sharpe"),
 )

 model.fit(X)
 model.weights_

This enables a whole new layer of user-defined and trustless indexes, based on advanced and computationally intensive financial machine learning rulesets.

Cards

Trust Guarantees​

Every step of this process is verifiable and reproducible:

  • Provenance: Methodology + data pinned on IPFS and CID pinned onchain;
  • Reproducibility: Anyone can rerun the pipeline and validate the output;
  • Audit Trail: All updates and rebalances are logged transparently onchain.

Active Management: ML-Optimized Vaults​

For users seeking more sophisticated exposure, Orion supports financial machine learning (FML) pipelines that inform allocation, rebalancing, and dynamic optimization.

Our approach draws inspiration from academic and institutional-grade tools such as:

  • Hierarchical risk parity;
  • Adaptive expected return models;
  • Online learning for portfolio allocation.

These models are either:

  • Hardcoded into the vault strategy, based on parameterized ML outputs, or
  • Queried dynamically via oracles or offchain computation layers (e.g., zkML, verifiable compute).

Orion abstracts this complexity for users, providing data, sdks and toolings, while allowing vault managers to go deeper — tuning model parameters, overriding logic, or composing hybrid strategies.

Agent