MGAT Alpha v3.1: could rule-based trading programs beat simply holding the same coins?
Eight automated programs traded cryptocurrencies on paper, every hour, for eleven weeks, against rules for judging that were written and signed before the first trade. This is the final account of what they did and what it shows.
No strategy family showed a reliable advantage over holding the same coins after fees. The programs lagged holding while prices rose and lost less while prices fell, which is what holding cash does, not what good trade selection does. The experiment was stopped on 9 November 2026 and will not continue with real money.
- Formal verdict: inconclusive for all four families. The outside price check that was meant to confirm every trade confirmed none of them, so under the signed rules no family could pass or fail.
- Picking small coins lost. The two small-coin families (BC, MED) lost money in the falling segment and had the largest samples.
- The AI adviser could not tell winners from losers. Trades it approved and trades it rejected lost almost the same (−5.3% and −5.5%).
- Three independent AI readers, from three different companies, read the frozen records and all recommended ending the experiment.
1The experiment
MGAT tested one question: does an automated trading program earn more than an investor who buys the same coins and holds them? Version 3.1 removed everything except the trading logic. It traded on paper: real market prices, simulated money, no slippage. That made the rules the only thing being tested.
A dedicated computer ran one decision round every hour, day and night. It held $8,000 of paper money in eight programs of $1,000 each: one each for the L1, BC and MED families and five for DEX, one per blockchain network. Each purchase used 15% of a program's available cash (minimum $20, maximum $600), at most three positions per program, with exchange fees of 0.1% to 0.3% charged on every buy and sell.
| Family | What it bought | Take-profit | Stop-loss | Time limit |
|---|---|---|---|---|
| L1 | 20 of the largest coins (Bitcoin, Ether, Solana and others) when up 2% in 24 hours and above their 20-day average | +12% | −8% | 7 days |
| DEX | New tokens on decentralised exchanges with rising price and enough trading; five programs, but only the Sui and BNB Chain ones ever traded | +25% | −12% | 3 days |
| BC | Coins worth over $100 million, up 3% in 24 hours, with a guard against buying a sharp fall | +20% | −10% | 7 days |
| MED | Coins that had been heavily sold (14-day RSI of 35 or less) or had just closed back above their 20-day average | +15% | −10% | 14 days |
Two AI advisers recorded an opinion on trades without any power to change them: a simple built-in rule from the start, and from 13 October a large language model (Sonnet).
2Test period and how it was judged
The programs traded from 25 August to 8 November 2026. The operator stopped the run on 9 November at 20:26 UTC, at 139 of the 200 planned trades; the books were frozen and fingerprinted at that moment. The rules for judging were signed before the first trade, so the results could not be read selectively afterwards:
- The run ends at 200 closed trades, on 30 January 2027, or earlier by the operator.
- A software repair on 7 October splits the run into segment 1 (before) and segment 2 (after). Each family is judged on segment 2 alone and the two are never pooled.
- A family passes only with at least 30 trades, a positive average after fees, a positive average even without its two best trades, and a better yearly return than both the Nasdaq-100 and the S&P 500.
- Every trade must be confirmed by a separate, outside price record.
Reaching 200 trades would have taken about five more weeks, and no family was close enough to the pass line for those trades to change the answer, so the run was stopped early.
3Results
The fleet as a whole
The 139 closed trades booked +$170.52. Four of them, on BNB Chain, took their prices from a single source that could not be cross-checked; one alone booked +$515.83. The signed rules exclude them. The 135 retained trades lost $227.15, an average of −1.4% per trade after fees, with 34% of trades profitable. At the stop the account stood at $8,154.90, including 7 open positions bought for $728.11. The largest fall from a peak was 5.4% as booked, 11.2% without the four unverifiable trades.
By strategy family
| Family | Period | Trades | Win rate | Average per trade | Without best 2 | Profit / loss |
|---|---|---|---|---|---|---|
| L1 | Segment 1 · 25 Aug–7 Oct | 17 | 53% | +2.8% | +1.5% | +$55.65 |
| L1 | Segment 2 · 7 Oct–8 Nov | 5 | 0% | −8.3% | −10.2% | −$58.45 |
| DEX | Segment 1 · 25 Aug–7 Oct | 6 | 67% | +9.0% | +2.6% | +$76.21 |
| DEX | Segment 2 · 7 Oct–8 Nov | 2 | 0% | −5.8% | — | −$18.61 |
| BC | Segment 1 · 25 Aug–7 Oct | 40 | 45% | +1.4% | +0.1% | +$49.33 |
| BC | Segment 2 · 7 Oct–8 Nov | 32 | 19% | −4.9% | −6.8% | −$179.19 |
| MED | Segment 1 · 25 Aug–7 Oct | 18 | 33% | −2.0% | −5.4% | −$49.22 |
| MED | Segment 2 · 7 Oct–8 Nov | 15 | 20% | −6.4% | −9.8% | −$102.87 |
| All | 25 Aug–8 Nov | 135 | 34% | −1.4% | −1.9% | −$227.15 |
In segment 2, the period the rules judge, only BC reached 30 trades, and it lost 4.9% per trade. L1 (5 trades), DEX (2) and MED (15) never reached the minimum.
Against simply holding the same coins
The pattern is the same in every family, which points to its cause: a program that is out of the market much of the time earns less when prices rise and loses less when they fall. Some of that idle cash was built in, since three of the five DEX programs ($3,000 of the $8,000) never made a trade.
How trades ended
The AI advisers
The built-in adviser approved every trade in three families and rejected every trade in the fourth, so it was a label for the family rather than a judgement. The Sonnet adviser was scored on 50 trades: the 5 it approved lost 5.3% on average and the 45 it rejected lost 5.5%. A permutation test gives p ≈ 0.97, meaning the difference is easily explained by chance. Its questions contained only information available at the time of each trade.
4Limits of the evidence
- No outside confirmation. The separate price record stored only the latest price, so it confirmed none of the 139 trades. Every figure is the program's own record.
- Small samples. L1 completed 22 trades and DEX 8. A small genuine advantage could hide in numbers this small.
- One market cycle. The period held one rise, to early October, and one fall. A sideways market was never tested.
- Paper is not real. Fees were modelled, but slippage, minimum order sizes and failed orders were not. Real execution would only lower the results.
5Conclusion
On eleven weeks of hourly decisions and 135 retained trades, no MGAT strategy beat holding the same coins. Selecting small coins lost money, the AI adviser had no predictive value, and the one useful behaviour, sitting in cash while prices fell, is a known property of trend-following rules rather than a discovery. The experiment ended on 9 November 2026 and was not moved to real money.
What remains is the method: rules for judging written before the data, every number re-checked by independent readers, and a plain account of what the data cannot show.