---
name: system-benchmark
description: Benchmark (4 Oct 2026) of 112 classic indicator trading systems (SuperTrend, Ichimoku, MA crosses, MACD, RSI, Bollinger, Donchian...) on 639 Robinhood Chain + Solana coins, 1h and 4h, Jul to Oct 2026; results, filters tested, caveats. Read before building or trusting any indicator based entry or alert.
sources: [cowork]
aliases: [system benchmark, top 100 systems, indicator benchmark, supertrend test, ichimoku test]
---
- [stated] 4 Oct 2026: Ben asked for an unbiased test of the top ~100 trading systems (SuperTrend, Ichimoku, MA combos with trend filters) on 1h and 4h, 300+ RH and 300+ Solana coins, last 3 months, ignore take profits, then 50 improvement ideas and execute the best.
## Setup
- Window 4 Jul to 4 Oct 2026 07:00 Bangkok. 311 RH + 328 Solana coins (GeckoTerminal pool lists by volume and tx count across many DEXes plus pump.fun graduated list; random sample; coin needs 400h+ of candles). Long only, signal candle close fills, 1% per side, no TP. First 100h per coin skipped; bad prints cleaned; 10 broken feeds and 92 low vol coins (stables, bridged majors, stock tokens) dropped.
- Score: geometric average result per coin ("typical coin"), plus edge vs random entries with same trade count and lengths on the same coin; time halves and coin halves.
## Results
- Only 2 of 112 above zero after costs. Best: BB(20,2) lower + RSI14<30, exit close > EMA20: +3.4% typical coin (RH 1h +2.6, RH 4h +0.7, SOL 1h +6.1, SOL 4h +4.4), win 55%. 2nd: Bollinger lower band, exit mid: +1.1%.
- Root cause: 71 to 78% of coins have negative lag 1 autocorrelation at 1h and 4h. Trend systems lose even with zero fees: SuperTrend 10,3 minus 33% gross, minus 51% net; Ichimoku TK cross minus 37%; EMA 9/21 + EMA200 minus 47%; MACD + EMA200 minus 56%; PSAR minus 63%. Golden cross 50/200 least bad trend system (minus 12%).
- Holds in dying coins (dip buy minus 6% vs hold minus 58%, SuperTrend 1h minus 79%) and healthy coins, so not a pure survivor effect.
- Dip buy family robust: 36 of 36 BB x RSI x exit variants positive; best BB2 + RSI<35, exit > EMA50: +6.3%, win 59%.
- One candle late entry kills it (typical coin about 0%): needs a bot or instant alert.
- Best filter: chain basket (equal weight all chain coins) above its EMA (200 1h / 50 4h): +3.7 pts per trade, win about 62%, all 4 cells positive, fewer trades. No help: 4h trend filter, SOL/ETH regime, age, ATR band, RS vs basket, stops (-25%, 3 ATR), time stop, ensembles. $20k volume floor hurt (edge partly in thin coins).
- Daily rotation into 72h gainers loses; into 72h losers shows huge paper gains driven by extreme rebounds and likely survivor bias, not trusted.
- Caveat: universe = coins still trading 4 Oct; next step is a point in time universe from chain logs.
- Files: ranking_112_systems.csv, ideas_50.md, code_and_results.tgz in /mnt/user-data/outputs/system-benchmark/ of Cowork session 97b5b89e (not public).
## v2, 4 Oct 2026: 6 timeframes and 72h losers deep check
- [stated] Ben asked for 15m, 1h, 4h, 8h, 12h and 1d, 500 RH coins (newer, so low timeframes), Solana and Base for higher timeframes, trades in the last 3 months, and a deeper check of the 72h losers rotation.
- Data: GMGN token_kline on the droplet pages with `to` in milliseconds (0.3s per call, far faster than GeckoTerminal); 4h/8h built from 1h, 12h/1d native. Cells: RH 15m 432 coins, 1h 551, 4h 631, 8h 594, 12h 622, 1d 466; SOL 335 to 376; Base 209 to 231 (1,310 RH coins sampled, many had no real 15m trading).
- No system is above 0 averaged over all 16 cells. RH: everything loses (hold minus 26 to 34% typical coin); 15m worst (fees). SOL: dip buy (RSI14 <30 / >70, BB lower + RSI) +2 to +5% on 1h to 12h; 1h version dies 1 candle late. SOL 1d SuperTrend 7,3 +11%.
- Base (older coins) is the exception: trend works on 4h to 1d. SuperTrend 7,3 4h +30% (bootstrap low +9%, both time and coin halves positive, still +31% one candle late), MACD > signal 8h +27%, TEMA 10/30 12h +32%, Stoch <20 dip buy 1d +41%. Universe is current survivors, so treat as TESTING.
- Autocorrelation: negative in 56 to 82% of coins on RH and SOL at every timeframe (weakest at 15m); Base near zero at 12h and 1d. That's why trend works only on Base.
- 72h losers rotation: the 8,800x (RH) came from survivorship (random 5 coins also made +247% on RH), a few extreme rebounds (cap picks at +50%: RH minus 76%), and dead-pool prints. Survivor-free test on 507 memory-screened coins (first mention date, 177 now dead, 28 Aug to 4 Oct): losers top 5 minus 42% vs random minus 75% vs winners minus 99.8%. Real as a relative effect (losers fall less), not a money maker. KILLED as a strategy.
- Files: ranking_v2_6_timeframes.csv, raw_v2_all_cells.csv and autocorr_v2.csv in /mnt/user-data/outputs/system-benchmark/ of this session. Candle bundles on the droplet: /opt/launchwatch/pub/hx-c9ois/tsb/ (mtf1.tgz, pit.tgz, m15more.tgz).