Walk forward
Sequential windows avoid future information leaking backward.
Projects / AXIOM Portfolio Intelligence
Versioned release2026-09AI-assisted portfolio research system
AXIOM Portfolio Intelligence
A portfolio research system for evaluating constrained strategies and producing source-grounded explanations of their results.
Problem
One-shot historical optimization and uncited AI commentary can hide leakage, turnover, cost and weak evidence.
Ownership
Designed the optimization and evaluation pipeline, FinBERT sentiment, FAISS retrieval, grounded commentary, interface, tests, container hardening and AWS path.
Primary proof
Equal weighting outperformed the optimized strategy on return and Sharpe in the documented walk-forward evaluation
Temporary external research endpoint; not a production financial service.
01 / Overview
System and intended use
A research reference for teams building platforms that combine quantitative analytics, document retrieval and source-grounded AI explanations—not an investment service.
84Automated tests
v1.0.0Release
Walk-forwardEvaluation
Equal weightStronger baseline
02 / Architecture
System boundary and flow
System flow06 stages
Market + document dataResearch inputs
Validation + featuresFeature pipeline
Walk-forward optimizerSequential evaluation
Risk + performanceRisk accounting
FAISS evidence retrievalGrounded retrieval
Research interfaceDecision surface
Evaluation evidence
Verified quantitative evaluation
Price-only walk-forward evaluation from 4 January 2021–31 December 2025 across AAPL, MSFT, GOOGL, AMZN and META, using a 252-trading-day lookback, monthly rebalancing, 2–35% asset bounds, a 5% annual risk-free rate and 15 bps transaction costs.
| Metric | Quantitative strategy | Equal weight | S&P 500 |
|---|---|---|---|
| Net CAGR | 16.83% | 20.59% | 13.12% |
| Sharpe ratio | 0.536 | 0.653 | 0.526 |
| Maximum drawdown | −39.63% | −46.55% | −25.43% |
Observed result: Equal weighting led on return and Sharpe ratio. The quantitative strategy reduced drawdown versus equal weight, but higher turnover created meaningful cost drag.
AI boundary: The optimizer produces portfolio weights. Retrieval supplies supporting documents and the language model explains the result using that evidence; it does not alter portfolio weights.
03 / Engineering decisions
Key design choices
Include friction
Turnover and costs remain visible.
Separate calculation from explanation
The optimizer determines weights; generated commentary can explain results but cannot change allocations.
04 / Verification
Claims, evidence and boundaries
| Claim | Inspectable Evidence | Boundary |
|---|---|---|
| The research system has automated verification. | The supplied project evidence records 84 automated tests. | Passing tests do not establish future investment performance. |
| The project has a versioned release. | The repository exposes the supplied v1.0.0 release link. | A versioned research release is not a production financial service. |
| Evaluation is sequential and keeps implementation friction visible. | The project records walk-forward evaluation with cost and turnover visibility. | Historical backtests do not establish future results. |
| The delivery path uses a hardened container and documented AWS workflow. | The project records a non-root Docker image and AWS OIDC/ECR/Systems Manager/EC2 delivery. | The supplied HTTP research endpoint is temporary and is not a production financial service. |
05 / Boundaries
Limitations and claim boundaries
- Research software and not financial advice.
- Equal weight outperformed optimization in the documented universe.
- Backtests do not establish future performance.
- The supplied HTTP research endpoint is temporary and may be replaced.
06 / Transferable capability
Related client problems
The system demonstrates how to combine deterministic analytical computation with document retrieval and source-grounded explanations while keeping model output outside the decision path.
Relevant to: quantitative research platforms combining analytical computation, document retrieval and source-grounded AI explanations.
Discuss a Similar Project →07 / Technical depth
Engineering appendix
Review implementation, failure modes, delivery and trade-offs
Implementation
- Implemented constrained optimization, sequential evaluation, FinBERT sentiment, FAISS retrieval and source-grounded commentary.
- Added an interactive research UI, automated tests, a non-root container and an AWS OIDC/ECR/Systems Manager/EC2 delivery workflow.
Failure modes and controls
- Validation rejects missing observations.
- Constraints bound concentration and infeasibility.
- Generated commentary never replaces quantitative output.
Delivery and operations
- Non-root container image
- GitHub OIDC to ECR and Systems Manager/EC2
- Reproducible v1.0.0 release
Engineering trade-offs
- Grounded commentary improves traceability but remains limited by retrieved sources and model behavior.
- Constrained optimization makes assumptions explicit while retaining model and historical-data risk.
Discuss a similar project
Need a similar AI/ML system?
Share the problem, current system, data or model constraints and the delivery environment you have in mind.