I connected Codex to a trading API and stress-tested an NVDA strategy

Follow this worked strategy

I connected Codex to the NexusTrade trading API and had it research NVDA, freeze intraday momentum rules, and inspect a five-fold walk-forward study with fee reruns and a trade audit. The agent operated this workflow through NexusTrade's MCP server; it did not build the integration. Native selection returned 14.13% against continuous NVDA at 49.20%, and one replay discrepancy stays open.

Follow the full journey

codex mcp add nexustrade --url https://nexustrade.io/api/mcp

codex mcp login nexustrade

Follow the Codex session

codex mcp add nexustrade --url https://nexustrade.io/api/mcp

This guide runs on NexusTrade's MCP server. Codex operated the trading workflow below, from connection to the live review. The agent's role here is operating the workflow through that server.

The table names the tool call, the approval checkpoint, and the artifact to keep at each stage. Placeholders such as a returned draft ID stand for values your own run returns. This page records no IDs from anyone else's run.

The recorded anchor is the walk-forward study, the fee reruns, and the trade audit below, including the unresolved first-fold replay difference.

Where Codex stops for approval

Each consequential action in this guide waits at an approval gate. The agent previews first and acts only after explicit approval.

The rules preview returns valid and issues before anything is saved. The agent resolves problems and re-previews instead of saving a candidate with open problems.

1. Approve the rules before the agent saves them

   The agent previews the configuration with build_portfolio and shows entry, sizing, exits, valid, and issues. It saves with create_portfolio only after approval and reads the returned draft back.

2. Approve the test cost before the agent submits once

   Historical tests are paid runs. The agent shows the exact request and estimated cost for approval before submitting once, then polls the same returned test ID until complete.

3. Approve the paper plan and verify it is paper

   The paper deployment is a separate portfolio with its own ID. The agent shows the plan for approval, verifies the deployment type, and leaves brokerage accounts unchanged.

4. Confirm the live move explicitly, or nothing moves

   The agent reads the actual account and broker capabilities first and keeps live trading unchanged until explicit confirmation. It reports an unsupported instrument or action and stops.

The nine-month lead and its replay discrepancy

The native optimizer reported 16.20% across the first three test blocks. Separate ordinary backtests of the selected rules reported 16.38%, versus continuous NVDA's 14.17%. The last two folds matched exactly, including all 77 daily NAV samples in each. The first fold differed by 0.15821 percentage points despite the same 28 round trips. Its cause remains unresolved.

The chart below uses ordinary replay NAV. Its 230 matched observations gave 5.08% sampled drawdown versus 21.25% for NVDA. The longer chart uses the native five-fold path. Keeping the paths separate avoids silently substituting the more favorable replay result into the native study.

The comparison curve is continuous NVDA rather than SPY because the candidate trades NVDA and its inverse ETF. Keep that benchmark scope attached to the return and drawdown figures.

Higher costs erased the limited-period lead

The same selected replay portfolios were rerun at higher fees without another optimization. They retained the same round-trip counts. At 6 basis points per fill, the compounded return was 14.95%, still just above NVDA's 14.17%. At 10 basis points it fell to 9.40%.

These are actual fee reruns on submitted notional, not a deduction estimated from trade counts. Next-bar price changes are already reflected in modeled fills. A larger flat fee does not establish quote-based spread and market-impact realism. The NVDA benchmark is a continuous holding-price path without repeated strategy trading fees; this comparison does not separately certify dividend total returns.

How the walk-forward selection worked

Training expands from January 2, 2025. Each fold has an earlier training window, a five-day embargo, a validation window and then a 91-calendar-day test block. The exhaustive grid has 36 settings: rolling means of 5, 20, 50 or 100 calendar days; inverse stops of 1%, 2% or 4%; and inverse profit targets of 2%, 4% or 6%. The other entry rules, long exits, sizing and fees stay fixed. Validation percent return selects the winner before its next test block.

The five-fold continuation preserves the first three fold calendars. The first three winners use a five-calendar-day mean and 2% inverse profit target, with inverse stops of 2%, 1% and 2%. The last two use means of 5 and 50 calendar days, 4% inverse stops and 2% profit targets.

Chronological splitting limits within-study lookahead, but the researchers inspected these historical dates repeatedly while changing strategy families. All reported results are retrospective development evidence. They are not a fresh independent out-of-sample test. A future test would require freezing both the rules and the selection procedure before observing new data.

What the trade audit established

The three ordinary replays produced 212 filled orders and 106 closed round trips. The audit reconciled cash, quantities and fees, found no direct stock shorts or simultaneous NVDA/NVD positions, and confirmed that all audited positions closed each day. Entries met the time limits. Eighteen sampled signal events had expected trigger values and no indicator errors; the full signal-causality history was not audited.

A representative NVD sale was about 763 shares. Twelve surrounding minute bars showed 1.45 million shares traded and no zero-volume bars, but the probe returned no bid/ask quote. That checks activity near one order, not the spread or impact of every fill. The first-fold replay difference remains open. The two additional native test folds were not independently certified with an exhaustive order audit.

Inspect the session and execution settings

Continue exploring