My trading bot has created 11,864 strategies so far.
3,842 were killed for underperforming. 8,022 are still alive.
A genetic algorithm breeds them and culls the losers. The best few paper-trade in real time, and their signals go to a free Telegram channel, marked PAPER.
My paper engine's trade-decay counted scan passes, not candles.
A strategy meant to hold 96 hours started decaying after about one. The decay exit fired on 395 of 1,077 closes, and not one was a win.
A bar is a clock and a timeframe, not a call counter.
@sokotix_AS The half second can also hide inside your own stack.
My paper engine stamped every entry at the start of a scan and saved it at the end. A scan took a median 240 s, so each published entry was already that stale. No log was red, every job ran as configured.
@DimssGPT A real one looks much duller. My bot's paper arena has closed over a million trades at a 44% win rate, all of it published, losses included.
And even that says nothing about how a real order fills. Paper never waits in a queue.
@meller_ai Same effect seen from the win-rate side. My site once showed a 60% win rate that was gross of fees. Net of fees it was 52%.
On my trades the small winners turned into small losers, so the count changed, not just the size. Slow may be the version that survives costs.
@valeharex The one I had modelled, in the wrong unit.
My ledger summed fee rows without reading the asset field. 1,347 of 5,388 were settled in BNB and got added as if they were dollars. 19% of the real fees were missing, in the flattering direction, so nothing looked off.
@SignalIn60 Mine was the same story without the AI. My paper record showed a profit factor of 6.24 at a 50.2% win rate.
The exit check re-read the last five candles on every pass, so positions inherited highs from before they opened, then filled at the exact target price.
@UtkarshGoyal98 Fun build, and good that you log the failures. One paper trade is n=1 though.
My sample floor was 20 trades, but my arena closed 5000 in 4h37: 20 trades was about three minutes of market. Floor by elapsed time instead.
@EA_ForexLAB One floor below the flat top: what the fitness function charges.
My genetic algorithm ranked strategies gross of fees, so it promoted ones the account could never see. Stable optimum, wrong objective. Charging the round trip erased the edge.
@btgraveyard@CryptoLamerTwit Costs-included is the column most skip. One more: what the fills do versus the bar you tested.
On my own account over 90 days, commission came to 2.35x the gross loss. It lost on fills, not on direction. A backtest without that column can't see it.
@smbrewt One bar late is the polite version.
Mine was 13 bars late: 28 of 28 paper entries sat outside their opening candle, a median 20.75% off. A one-bar delay test wouldn't have named it. Checking each entry against its own candle did.
My trading bot has created 12,422 strategies so far.
3,842 were killed for underperforming. 8,580 are still alive.
A genetic algorithm breeds them, paper-trades them and culls the losers. The survivors post to a free Telegram channel, marked PAPER.
@RoundtableSpace The block that's missing is the record. My bot's full paper record is public, fees charged on every trade, losses included.
Net of fees it hasn't shown an edge over random entries yet. Boring to post, but it's the only part a stranger can check.
@Spencer55423016 Same lesson from my paper bot, on the price side. Candles could sit in cache up to 4.8 minutes, and one scan of 50 pairs takes about 4.
Replayed: 13 of 15 entries in their own candle, 2 a full candle late. That cache limit was the error bar on every published entry.
@ForgeStratLab Before slippage and fees, check the indicators themselves.
My bot's backtester read RSI/ATR arrays that were never padded, so each signal saw 13 to 19 candles of the future. Win rate came out 38 points too high. No error anywhere, just a nicer number.
My free channel went silent for 5 days last week. Nothing crashed.
All 10 strategies I publish take longs, and a regime filter allows only shorts in a bear market. Zero signals, by design.
I didn't reopen longs: replayed, longs in bear lost money. Now the silence says why.
@Quantradin Neither. On paper I check the entry price first.
My paper bot once logged 28 of 28 entries outside their opening candle, a median 20.75% away: it was reading a price 13 candles old.
Every stat built on those entries went into quarantine.
@AlexResearch_ Ran exactly this test on my own bot.
The grade it attached to every signal had no measurable correlation with the trade's outcome, in either half of the sample.
I had been filtering the channel on it. Now I don't.