A viral AI trading repo says one sentence in a terminal can replace a quant desk, and its own example made +36% on Bitcoin in 2024. Bitcoin itself made +111% that year. We rebuilt the 20/50 moving average rule in Trigr, tested it from 2020 with fees, added a 200-day filter and ran it on Ethereum. Verdict: fade the one-year example, trade the filtered version.
The viral claim
The repo, Vibe-Trading, passed 36,000 stars on GitHub. The post that made it go viral says hedge funds pay quant PhDs six figures to find alpha, and this repo turns the whole quant desk into one sentence you type in a terminal. Its showcase prompt asks for a backtest of a BTC 20/50 moving average strategy for 2024.
The repo's own answer was +36% for 2024. That sounds fine until you compare it with just holding Bitcoin, which made +111% that year. A strategy that trails holding by 75 points in a strong year is not a quant desk. But one year is also too short to judge anything, so we tested it properly.
See the original post on X and the Vibe-Trading repo on GitHub.
How we rebuilt it in Trigr
We described the rule in plain English and Trigr turned it into fixed rules: no AI guessing trades, only math on closed daily candles. Every run includes trading fees and slippage.
| Setting | Original rule | Upgraded rule |
|---|---|---|
| Market | Bitcoin (BTC) | Bitcoin, then Ethereum |
| Timeframe | Daily candles | Daily candles |
| Buy | 20-day average above 50-day average | Same, and price above its 200-day average |
| Sell | 20-day drops below 50-day | Either condition fails |
| Direction | Long only, no leverage | Long only, no leverage |
| Test window | Jan 2020 to Oct 2026 | Jan 2020 to Oct 2026 |
| Costs | Fees and slippage included | Fees and slippage included |
The original 20/50 rule since 2020
Over almost seven years, the original rule made +646% on Bitcoin. Honestly, not bad. The problem is the worst moment: a -59% drop. At one point $10,000 at the top became about $4,100 at the bottom. Most people quit right there.
It also shows why one-year backtests fool you. In 2024 alone the rule lagged Bitcoin badly, which is exactly the year the repo picked for its example.
The fix: one 200-day rule
To fix the drop we added one rule Paul Tudor Jones is famous for: when price is below its 200-day average, you are out. The rule now needs the 20-day above the 50-day and price above the 200-day.
| Bitcoin, Jan 2020 to Oct 2026 | Return | Worst drop |
|---|---|---|
| Original 20/50 | +646% | -59% |
| 20/50 + 200-day filter | +639% | -41% |
The return barely moved, but the worst drop went from 59% to 41%. That is the difference between a strategy people quit and one they might actually keep running.
We also kept the last part of the data aside while building. From February 2025 to today, the filtered version made about +30%, while just holding Bitcoin was down about 19% over the same window.
Then Ethereum
The same filtered rule on Ethereum made +2,487%. Just holding Ethereum made +1,623% and fell 79% on the way. The upgraded version's worst drop was 50%. So on ETH it beat holding and cut the worst drop.
We tried to break it
A result like this can be luck, so we moved every setting about 20% up and down: faster averages, slower ones, a different trend filter and higher costs.
| Variant | Return | Worst drop |
|---|---|---|
| 20/50 + 200-day (the upgrade) | +639% | -41% |
| 16/40 + 200-day (faster) | +888% | -32% |
| 24/60 + 200-day (slower) | +1,004% | -39% |
| 20/50 + 160-day filter | +634% | -37% |
| 20/50 + 240-day filter | +621% | -40% |
| Higher costs (5 bp slippage) | +628% | -41% |
Every version stayed above +600%, and the worst drop stayed between 32% and 41%. A strategy that only works with one exact setting is a warning sign. This one did not fall apart when we pushed it.
Whose strategy is it?
This is not Paul Tudor Jones's strategy. The entry and exit are the repo's 20/50 crossover. His 200-day rule is only used as a filter on top.
Verdict: trade it or fade it?
- The one-year 2024 example: fade it. It trailed holding by a wide margin and one year says very little.
- The 20/50 crossover since 2020 with the 200-day filter: trade it, with care. Similar return, a much smaller worst drop, and it held up when we moved the settings.
Why one-year backtests mislead
The repo's example is a good reminder of a common trap. A single year can make almost any rule look good or bad, depending on whether that year trended, chopped or crashed. 2024 was a strong year for Bitcoin, and a slow crossover always lags in a fast rally, so the rule looked weak next to holding. Run the same rule from 2020 and it looks much better, because it also includes the 2022 bear market, where stepping aside saved a lot.
Neither view is the full truth on its own. What you want is a long window that includes bull and bear markets, real fees, a held-back period you never tuned on, and a check that nearby settings give similar results. That is what we did here, and it is why the verdict is different for the one-year example and for the longer test.
Try it yourself
This is a research article: the upgraded rule is not listed as a strategy. You can rebuild it in Trigr in plain English and backtest it with fees. Before you trust any backtest, read how to spot overfitting and why live results differ from backtests, then paper trade it first. The Trigr docs explain how fees, slippage and the held-back data work.
Sources: the Vibe-Trading repo on GitHub, the viral post on X, and Stockopedia on the 200-day moving average.
Backtest, not live. BTC and ETH daily, January 2020 to October 2026, fees and slippage included. Past results are not a promise of future results, and this is not financial advice.