How the test was run
- Data. Binance daily candles from January 2019 for every USDT pair ever listed, delisted ones included. Series are split where a symbol was reused after a pause or redenominated (a day that opens 20× away from the last close).
- Weekly lock. At each Monday 00:00 UTC close from 2019-07-21: the 60 coins with the most volume (at least 200 days of history, $2M a day), ranked on nine factors with that week's learned weights; the top 5 that close above their 200-day average are the picks.
- Levels and exits (v2). Stop = 1.5 × volatility × √days (8-50% below), counted on daily closes; take-profit = 2× the stop distance, a resting order; otherwise the close on the end date; 0.1% fee per side and 0.02% slippage.
- Learning without hindsight. The weights used on a Monday came only from weeks whose holding period had ended by then (rank correlation of each factor with every universe coin's result, 52-week half-life, at most 70% from the record). The live model continues the same history.
- Only finished weeks count. Weeks whose holding period has not ended are left out of every statistic, so recent picks that hit a level early cannot tilt the numbers.
What the result means
Picking the strongest established coins and holding them made money on average from 2 months up, but it did not beat the simplest alternative, holding Bitcoin, in this period. Two things explain most of it: from 2019-2026 the majority of altcoins lost ground to BTC, and stops closed picks near their lows (v2's wider close-only stop helped, but did not close the gap). Without any stop the 6-month and 1-year picks beat BTC on average, but only through a handful of giant winners, with most picks still behind BTC and some going to zero. The learning reacted to this by favouring large, calm, long-listed coins, which is why BTC, ETH and BNB are the most-picked coins.
Limits: the factor list, the starting weights and the level rules were set before the test and not tuned, which keeps the result honest but also means nothing here was optimised. Weekly picks overlap, so the effective number of independent results is much smaller than the pick count. The future can look different from 2019-2026. Compare with the Trend Breakout setup and the other backtests.