Skip to content

methodology.md

Latest commit

 

History

History
113 lines (91 loc) · 4.67 KB

methodology.md

File metadata and controls

113 lines (91 loc) · 4.67 KB
Raw

FusionFinance methodology

Controlled comparison contract

The next claim-valid three-arm run must use configs/fusionfinance-demo.json without changing assumptions after seeing the result.

Assumption Locked value
Window 2026-02-02 through 2026-07-09
Starting capital $10,000
Benchmark SPY
Rebalance cadence Every 10 sessions
Execution Next session open
Transaction cost 5 basis points of traded notional
Slippage 2 basis points of traded notional
Maximum gross leverage 1.0x
Maximum absolute position 10%
Annualization 252 sessions
Annual risk-free rate 0%

The locked 16-name universe is AAPL, AMZN, AVGO, BRK-B, COST, GOOGL, HD, JPM, LLY, META, MSFT, NFLX, NVDA, ORCL, TSLA, and WMT. SPY is benchmark-only and cannot be proposed as a position.

Strategy arms

The controlled experiment requires three proposal generators feeding the same execution function:

  1. Pure ML uses structured quantitative signals and no LLM recommendation.
  2. Pure LLM uses the LLM recommendation and no proprietary ML signal or verifier forecast as an input.
  3. FusionFinance begins with a proposal, audits its evidence, challenges it with the independent structured verifier, and permits only policy-approved target weights to reach execution.

The implementation must not average arms or give FusionFinance a different cost, timing, or leverage model. Empty or rejected proposals remain cash.

Timing and look-ahead controls

  • Market features are built from rows on or before the decision timestamp.
  • Forward labels used to train the verifier begin strictly after their feature timestamp.
  • Filing availability is derived from an explicit timezone-aware acceptance timestamp and the observed exchange-session calendar.
  • Decisions are associated with a close and execute one full session later at the next open.
  • Proposals outside the shared rebalance clock, outside the universe, after the final executable session, or above risk limits fail rather than being clipped silently.
  • A thesis must be committed before verification. Evidence sources must already be available at the thesis timestamp.

The tests demonstrate that identical weights receive identical P&L across arms and that changing later market data cannot alter an earlier fill. These tests validate the kernel; they do not prove that a proposal generator itself is free of leakage.

Accounting and metrics

Positions are expressed as target weights of pre-trade portfolio value. The kernel trades at the next open, deducts transaction cost and slippage from cash, and marks holdings at each close. Portfolio return for session t is:

wealth[t] / wealth[t - 1] - 1

Total and annualized returns are compounded from wealth, not summed. Maximum drawdown is the worst wealth-to-running-peak ratio. Volatility uses sample standard deviation; Sharpe and Sortino use the locked risk-free rate. The ledger also records fills, turnover, total execution costs, cash, gross exposure, and net exposure.

When a same-length benchmark wealth sequence is supplied, the metrics layer additionally reports benchmark return, wealth-relative excess return, tracking error, information ratio, beta, and annualized alpha. Misaligned, non-finite, or non-positive benchmark wealth is rejected by length/value checks. Date-index alignment is currently the caller's responsibility and must be enforced inside the contract before a claim-bearing run.

Evidence levels

Artifact or result Current evidence level
Shared execution invariants Implemented and unit-tested
Wealth and benchmark metrics Implemented and unit-tested
Exact-citation, numeric, and timestamp audit Implemented and unit-tested
Filing transforms Implemented from an attributed subset and integration-tested
AMD training workload Semantically cross-checked self-reported receipts; hashes protect integrity, not independent attestation
Checked-in three-arm replay Provisional legacy visualization only

Independent review also requires proposal/input lineage, config and market-tape hashes, post-cost leverage enforcement, ledger/result reconciliation, strict schema coercion, and date-bound benchmark alignment before this foundation can produce comparative evidence.

Reproduction checks

pytest -q tests/test_fusion_experiment.py tests/test_fusion_evidence.py \
  tests/test_filing_alpha_integration.py
make artifacts
make verify
make test

build_fusion_artifacts.py reads only checked-in evidence, fails closed on missing fields, and deterministically preserves the legacy replay's provisional flags. It does not convert that replay into a controlled experiment.