TRADAR

Methodology

Tradar is a paper-trading experiment: 11 AI models each manage a ring-fenced R20,000 portfolio on the VALR exchange, trading ZAR-quoted crypto spot pairs. No real money is at stake.

Rules

• Each model receives the same research pack and decides independently, weekly.
• Per-order limit: 25% of portfolio value; min R100.
• Max 10 orders/week per model; max 2 taker.
• Drawdown freeze: a model is frozen to HOLD-only if its portfolio falls below 40% of start.

Fee model (VALR, verified 2026-07-26)

ZAR pairs: 0.36% maker / 0.70% taker round trip. Fees are applied identically to every model so comparison is honest.

Cadence economics

Median realised move on VALR: 1.28% daily, 6.70% weekly (100 periods, measured 2026-07-26). Weekly trading needs ~53% directional accuracy to beat fees; daily needs ~64%. Full-turnover fee drag is ~19%/yr weekly vs ~131%/yr daily.

Schedule (SAST = UTC+2, no DST)

• Monday 09:00: research pack built, all models decide, orders placed.
• Wednesday 09:00: unfilled post-only orders cancelled.
• Saturday 08:00: weekly snapshot frozen; this is the scoreboard of record.

Known limitations

• The experiment is young. A small field over a few weeks proves nothing statistically.
• The research pack is generated by LLMs and can be weak or repetitive; it is a separate workstream (D10).
• Paper fills simulate against the live VALR book but no real orders are placed.

The null hypothesis (IT3 R10)

Null: no model has skill; all differences are noise and market exposure.

Current verdict: null not rejected. Null not rejected: no model is statistically distinguishable from zero return at 95% confidence. Consistent with the prior art.

Evaluated continuously at each Saturday snapshot. A model rejects the null only if its excess-return confidence interval excludes zero at 95%. Below 4 observations the verdict is insufficient data.

Statistical honesty

Whether any model beats any other is not a success metric here. We publish this limitation deliberately. Credibility comes from reproducibility: every model sees the same inputs, and every JSON endpoint is open for independent verification.

Prior art

The prior art has not found LLM trading edge. Alpha Arena (nof1.ai) ran six frontier LLMs with real capital; results ranged from roughly +20% to −30% and −62%, with most models losing money across seasons. Factor-adjusted academic work finds headline returns are largely explained by passive market exposure rather than selection alpha. We cite this not to pre-judge our own result, but to set the honest prior: the most likely true answer is that the null holds.