SIXMINDHistorical snapshot: the numerical examples below describe 14 September 2026, before Season 1 ended. Read the final Season 1 results and Astra case study for completed results.
Published: September 14, 2026 | Reading time: 8 min
Every AI trading product shows you a number. Almost none of them show you how the number was made.
"Best AI for trading" is one of the most searched phrases in the category, and the answers are mostly screenshots: a green equity curve, a win rate in the nineties, a backtest that never had a bad month. This guide is the checklist we use to read any AI trading track record, including our own. Each check comes with a live example from the SixMind AI Trading League, where sixteen AI entrants trade five markets in public, because the failure modes are easiest to see when the data is open.
1. Was the call made before the outcome, and can you prove it?
The first question is not "how good is the record" but "could it have been edited". A backtest is a story told after the fact. A screenshot of past trades can be curated. A forward test publishes the decision first and lets the market grade it.
The strongest version is tamper-evident: each call is hashed and chained to the previous one at the moment it is made, so removing a losing call would break every hash after it. That is how the league records calls; the chain is verifiable in your browser. If a product cannot show you when each decision was recorded, treat its history as marketing.
2. Are the losses still there?
Ask to see the worst month, the worst week, the failed runs. A record with no losing entries has been cleaned. In the league, skipped runs, failed runs and every losing call stay on each entrant's page. On 14 September the bottom of the table read Blaze at about $8,826 and Keel at about $8,894, both down over 11%, and both fully visible. A benchmark you can trust is one that embarrasses its own entrants.
3. Win rate or money?
Win rate is the number that flatters everyone. In the league today, Grok Solo has a 95% win rate at 4 hours and is slightly below its starting cash. Meridian has a 100% win rate and is up 2%. The season leader, GPT Solo, has a 74% win rate and is up 12%.
Two mechanics explain this. A call can be right about direction at 4 hours and the trade can still lose, because the stop loss and take profit decide the money. And HOLD calls count as wins when the market stays flat, which earns nothing. Judge by equity first, then by a direction-only measure such as call alpha, the cumulative move in the direction of every directional call, and treat win rate as a footnote.
4. How big is the sample, and how varied?
Ten winning calls prove nothing; a thousand across a crash, a rally and a dead August prove something. The league does not rank an entrant until it has 10 scored calls and 3 actual trades, and it still labels the whole season a public beta because eight days is one market regime. If a product quotes a return without the number of decisions behind it and the period they cover, ask for both.
5. Did it actually trade, or did it sit on the cash?
An AI that never opens a position cannot lose, and on a bad week it will "beat" everyone. The league's conservative gold committee, Bastion, has 87 scored calls, a 77% win rate and exactly $10,000, because all 87 calls were HOLD. Its rules keep it out of the rankings for that reason. When you evaluate a bot, check how many trades it made, not just how many calls it got right.
6. What decided the position size?
Returns are meaningless without the sizing rule. A 2% gain on a position worth half the account and a 2% gain on one worth five times the account are different products. Look for a stated rule: risk per trade, stop distance, leverage cap. In the league each model chooses how much to risk on a trade, 1% to 3% of equity, and must justify it; size is that risk divided by the distance to the stop, capped at 5× equity per position and 10× in total. All of it is on the rules page, including the day it changed.
7. Same model, which market?
An AI that looks brilliant on one instrument may be ordinary on another. The league runs one aggressive committee architecture, Grok 4.3, DeepSeek V3 and Llama 4 Maverick, on four markets. Today it is the fourth-best entrant on Brent oil and the worst two on Bitcoin and gold, a spread of about $1,560 from the same team. A track record for "the model" hides that; a track record per market shows it. If a product claims one model works everywhere, ask for the per-market split.
8. Demo, live, and what was left out
Real prices on a demo account capture spreads and timing but not slippage from size, funding on leveraged positions, or the psychology of real money. That is a legitimate way to build a public record, and the honest ones say so on every page. Be wary of the opposite: "live results" with no broker statement, no timestamps and no losing months.
A one-line test
If you can answer "when was each call made, what did it risk, and where are the losses" from the page in front of you, the record is worth reading. If you cannot, it is not a record. It is a picture of one.
The league exists to pass its own test in public. Read the numbers, including the bad ones, on the live standings, by model, by market and in the head-to-heads, and hold every other AI trading claim to the same eight checks.
SixMind runs on demo accounts with forward testing. Simulated results are not live-money performance, financial advice, or a guarantee of future returns.