← Portfolio
Trading research · Oct–Nov 2026 · updated Nov 15 · all code + data public

What we found in four markets, tested honestly.

Thousands of rule versions across memecoins, crypto, stocks and forex. Every one was judged on data it was not chosen on: locked years, hidden markets, tests saved to git before they ran. This page sums up what failed and what held up. It doesn't say these markets can't be beaten, only that nothing we built beat them on this much data.

Memecoins
Every lane lost
2,639 paper trades in 16 runs, −$2,665, 0 runs profitable. 16,071-coin replay: every exit loses.
details →
Crypto (SOL, JUP, BONK, WIF)
0 of 436 beat holding
7 months of hourly prices. Random rules did as well.
details →
Stocks (US ETFs)
Nothing beat SPY
Locked year: best rule +4.9% vs just holding +12.9%.
details →
Forex
1 effect held up
~1,500 indicator versions failed hidden data. The Fed-day dollar drop held on 4 pairs.
details →

Memecoins on Solana

Trade Desk: a bot that scouts brand-new Solana coins, screens them with Solana Pulse safety checks, and paper-trades them with different exit rules.

LOST
No lane, exit rule or entry filter made money.Buying new memecoins from public scans loses by design: the coins that run 13×+ are the same minutes-old pools that rug.

What we tested

TestDataResult
6 live paper lanes, $1,000 eachMain, Fast, Ladder A/B, WAIT-5 momentum, wallet copyAll 6 lost: 641 trades, −$720 since the Nov 7 reset
Every paper run, Nov 1–1516 runs across all lanes; Ladder A kept running after the others were retired Nov 90 of 16 ended in profit: 2,639 trades, 909 wins (34%), −$2,665. Ladder A alone since Nov 9: 986 trades, −$581
Strict replay, every exit rule16,071 coins with 1-minute candles (Nov 1–4)Every exit loses: typical trade −11.6% to −17.9%
120-cell exit gridTake-profit 10–100% × stop 6–40% × 30–600 min, trailing stopsNo lead
Entry filtersSafety verdict, liquidity, volume, buy/sell ratio, pool ageSafer coins stop the rugs but still bleed ~2% a trade to costs
6 brand-new strategies (Claude + Gemini)Dip buys, breakouts, second-wave, capitulation bounce, flow; 5,705 coins at 1%/side0 leads; big samples −5% to −19% a trade
Copy the "smart wallets"100 coins, 105,857 transactions, every wallet's profitFailed: early winners lost on later coins (−12.6 SOL)

Who actually makes money on a new coin

From the wallet study (one row = one wallet trading one coin, inside the first 2 hours):

WhoWalletsProfitableTotal
Sold more than they bought (got tokens before trading opened)2,82895.8%+1,816 SOL
Pool creators4959.2%+124 SOL
Bought in the launch slot16734.7%−187 SOL
First buy within 10 s1,68518.7%−584 SOL
First buy 10–60 s8,0334.1%−6,417 SOL
First buy 1–10 min18,4101.1%−1,709 SOL
First buy after 10 min3820.5%−31 SOL

What we found out

The money goes to insiders.Wallets that held tokens before trading opened made +1,816 SOL. Every group that bought after launch lost: about 8,900 SOL combined. By the time a bot sees a coin, it is already late.
"Smart wallets" were mostly luck.51 wallets won on 3+ early coins. On later coins they won only 39% of the time and lost 12.6 SOL in total.
Stops don't stop rugs.Crash fills came in at −70% to −100%. Three rugs caused $224 of the first run's $238 in losses.
Spotting scams works; trading doesn't.Coins Pulse rated AVOID were 82% dead soon after, and 0 of 15 SAFE coins died. Telling rugs apart is possible. Profiting from new coins is a different problem.

full write-upraw data + scripts23,379 coins of candleslive paper desk

Crypto: established Solana tokens

Not new launches: timing rules on established tokens (SOL, JUP, BONK, WIF), plus a check of whether safety ratings match what happened next.

LOST
0 of 436 rules made money and beat simply holding in both periods.Random entry and exit times with the same amount of trading produced 2. Being out of a rising market costs more than timing saves.

What we tested

TestDataResult
436 timing rules (109 per token)~7 months of hourly price + volume, each token's main Solana pool (GeckoTerminal). Rules found on the first 60%, judged on the last 40%. Costs 0.1–0.3% a side0 beat holding in both periods (random rules: 2)
Rule familiesMoving-average crosses, price above average, Bollinger bounce and breakout, RSI, breakouts, momentum, volume spikes, hour of day, day of week, dip buys in uptrends25 made money in both periods. All 25 trailed holding, because the test period rose 19–35% on every token
Luck check vs random timingThe same number of trades at random times, many runs3 beat random in both periods. They are candidates for a forward test, not edges, because they were picked from 436

The 3 candidates (not proven): forward test, Nov 1–15

Same code and costs, run on the 14 days of prices that came after the study. Only 2–6 trades each, far too few to call an edge.

RuleTrainTestRandom did as wellForward (vs holding)
JUP: in an uptrend (above 168-h average), buy below the lower Bollinger band, sell at the middle band or after 48 h+11.0%+31.1%9% / 1%+3.2% vs +4.3%, 3 trades
JUP: after an hour with 5× normal volume and price up, hold 24 h+13.9%+40.7%15% / 2%+10.0% vs +4.3%, 2 trades
WIF: buy below the lower Bollinger band (20, 2.5), sell at the middle band or after 72 h+6.8%+15.9%4% / 8%−1.5% vs −25.2%, 6 trades

What we found out

In a rising market, holding wins.Every token rose 19–35% in the test period. A rule that sits in cash part of the time has to be very good to keep up.
Mean reversion was the only family that showed signs of life.Buying sharp dips below the Bollinger band on JUP and WIF beat random timing. That is still 3 hits out of 436 tries, so it needs a forward test.
Established tokens behave nothing like new launches.All 15 coins Pulse rated SAFE (WBTC, ETH, Bonk, HYPE, ZEC and others) were alive and just moved with the market. 82% of the AVOID coins died.
Seven months is short.One trend, one market mood. The stock and forex tests below used 5–20 years for that reason.

rules, luck test + fetch scriptsafety verdict check

Regular stocks (US ETFs)

Night Desk: an overnight desk that searches for stock rules and paper-trades only the ones that beat holding the S&P 500 (SPY) after costs.

LOST
No timing rule beat just holding SPY.An AI-assisted search found 3 that did beat SPY on the data they were picked from. On the locked final year they made a third of what holding made.

What we tested

RoundRulesResult
120-day momentum + breakout, 60-day momentum above the 50-day average, pullbacks in an uptrend. Walk-forward test with the newest 40% unseenBest +10% vs SPY +30%
2 (AI search, 27 rules)200-day trend filters, sector rotation, panic dip-buying, volatility timing. Searched on Jul 2020 → Nov 2025; the newest year was never written to diskFinalists beat SPY on search data (+67.6% vs +61.5%), then lost on the locked year
Live paper desk (Alpaca, Nov 9–15)The desk places only tiny "unproven" $50 probes until a rule passes; it found 2 (LLY Nov 9, COST Nov 13)+$1.81 open on $100 of probes. Far too small to mean anything; no rule has passed yet

The locked year (Nov 2025 → Nov 2026), tested once

StrategyReturnMax drawdown
Just hold SPY+12.9%8.9%
Trend + panic dip (20/200)+4.9%8.9%
Trend only+4.7%8.9%
Trend + panic dip (15/200)+4.0%8.9%

What we found out

Holding the index is a hard benchmark.Momentum, breakouts, pullbacks, trend filters and sector rotation either lost to SPY or only beat it on the data they were tuned on.
Trend filters trade return for comfort.They cut drawdowns in some years but cost more return than they saved.
Count your tries.With 27 rules tested, the adjusted confidence that the "winners" were real was only ~0.70, below the usual 0.95 bar. The locked year confirmed it was luck.
Two bugs made it look better than it was.A fake 88% drawdown came from one engine bug. A confidence check got yearly numbers where it needed daily ones. Both were caught and fixed.

Alpaca daily bars, 13 ETFs (SPY, QQQ + 11 sectors), Jul 2020 → Nov 2026, 5 bps a side · full write-upevery rule triedprice data

Forex

It started with an old tip: "a 5/16 moving-average crossover with Japanese candlesticks on EUR/USD is profitable." We tested that, everything people usually add to it, and then effects from published research.

MIXED
Indicator systems: no real edge. One effect from published research held up.The best EUR/USD system lost on 9 other pairs and on 2005–2015. Selling the dollar on Fed announcement days held on 4 pairs over 21 years.

Indicator systems (~1,500 versions)

TestedResult
5/16 SMA + EMA crossover, daily 1999–2026Made a little before 2015, lost since (−2 to −3%/yr)
Same on 15m / 1h / 4h / daily + candle patterns15m and 1h lose everywhere (15m: −19,795 pips 2016–21). 4h + engulfing was the only positive one
EMA ribbon 5/16/35/68 with volatility gates0 of 45 combinations positive in all 3 periods
"Everything" system + 1,260 stop/trail/target settings4h: ~70–76% of settings positive on EUR/USD. 1h and 15m: 0 of 1,260
Self-tuning version that "learns as it goes"Chased last period's winner and did not beat a fixed setting
No stop, take-profit onlyShowed "100% wins" while holding losers up to 5½ years at −2,000 pips. A trap
Hidden-data test (saved to git before it ran)Best 4h system, unchanged: −61.9R on 9 other pairs (515 trades), −630 pips on EUR/USD 2005–2015

Effects from published papers (rules fixed from the papers before running)

EffectResult
Fed day: sell USD from 5 PM New York the evening before a scheduled Fed announcement until 5 PM on the day
Mueller, Tahbaz-Salehi & Vedolin, Journal of Finance 2017
Holds. EUR/USD 2005–2026: 174 events, +2,717 pips, t = 2.3; also positive on GBP, JPY and AUD. The same trade on other days: −2 to −3 pips. Last 12 months negative.
Month-end rebalancing into the London 4 PM fix
Melvin & Prins 2015
Fails: EUR/USD −1,911 pips
A currency weakens during its own trading hours
Breedon & Ranaldo 2013
Real but not tradeable: +2.25 pips/day before costs (t = 4.8); the 1.5-pip spread eats it

What we found out

Costs decide short timeframes.On 15-minute and 1-hour charts the spread ate every rule. One rule paid ~14,800 pips in spread over 6 years.
Stops were the ingredient that mattered.On 4h charts the move-to-breakeven and trailing stop did more than any indicator.
An effect with a reason beats a chart pattern.The only thing that held up is a dollar flow with a known cause, not an indicator mix.
Mixing unrelated systems helps.Trend + Fed day raised return per drawdown from 0.41 to 0.63, about 16%/yr at a 25% max drawdown in a typical backtest year. All 4 systems now run as a forward paper test, started Nov 9, 2026. First week: 1 closed trade, +85.6 pips (4-hour trend A, Nov 10–13); no Fed-day trade yet (next meeting Dec 9). Paused Nov 15: the practice-account data feed needs a new key.

Oanda candles (15m → daily, 10 pairs, measured spreads, back to 2005), FRED, federalreserve.gov · full write-upfindingsevery price file

How we tested, and why most backtests lie

Our biggest early "wins" (+2,000% a trade, a 6,526× coin, Sharpe ratios above 3) were all bugs in the test, not edges in the market. These rules are how we caught them.

  1. Real data, named. Every price has a named source. Missing data is "unknown", never filled in.
  2. No look-ahead. A signal on a candle's close fills at the next candle's open. If a stop and a target hit in the same candle, the stop counts first.
  3. Real costs. Measured spreads per currency pair, 5 bps a side for ETFs, 1% a side and −100% write-offs for dead memecoins.
  4. Several periods, not one total. A rule has to make money in each period, not just overall.
  5. Locked data. The newest stock year was never written to disk during the search, and the script refuses a second run. Forex kept 9 other pairs and 2005–2015 hidden.
  6. Save the test before running it. Hidden-data tests were committed to git first, so nobody could adjust them after seeing results.
  7. Count the tries. Adjust confidence for how many rules were tested. For settings grids, report how much of the grid works, not the best cell.
  8. Compare with doing nothing clever. Stocks vs holding SPY, crypto vs holding the token, rules vs random timing.
  9. Then forward-test. The future is the only data none of these systems has seen.

methodhardships + how we got past themwhole repo