How the SIMMENT engine trades, and how it has actually done

Live track record (engine-only trades)

Hand-placed trades on the same account (excluded above): 4 positions, 11 exits, +5,655.9 USDT. Whole account: 13,459.66 USDT (+34.6%). Exit mix: stop loss 93 · tp1 74 · stop be 61 · tp2 28 · tp3 13 · close all 3. Open positions now: 3. Updated Tue, 13 Oct 2026 12:10:49 GMT. Numbers refresh every 10 minutes from the ledger; 0 historical fills were repriced after a data bug was fixed on 5 Sep 2026 and are labelled in the trade history.

Architecture in one paragraph

The engine has two layers. The decision layer runs in the desk page: every 30 seconds it takes one of seven pairs (BTC, ETH, BNB, ADA, SOL, AVAX, XRP), pulls 200 one-hour candles from Binance, computes RSI 14, MACD 12/26/9, Stochastic RSI, Bollinger 20/2, volume SMA 20 and an EMA 50 trend, and produces a confidence score. The position layer runs on the server and manages every open position with a stop loss, a three-step take-profit ladder, a break-even move and a trailing stop, priced from the server's own Binance feed.

The four gates before any fill

1. Technical score. EMA 50 trend +3, MACD +2 (+1 when the histogram is large), Stoch RSI +2 under 30 or over 70, Bollinger +1/+2 by band position, volume +1 above 1.5x average. A bullish share of 60% or more sets a bullish trend with confidence 50 to 95; 40% or less sets bearish; in between is sideways with confidence 55.
2. Sentiment, contrarian. Fear & Greed under 25 or over 75 adds 15 confidence points, under 40 or over 60 adds 10. News sentiment can move confidence by up to ±25. Extreme realised volatility subtracts 20.
3. LLM opinion. A Groq-hosted Llama 3.1 8B model receives price, 24h change, RSI, MACD histogram, volume and Fear & Greed and answers with a JSON signal and confidence. A BUY or SELL can only raise confidence to its own level; it is called at most every five minutes per pair and only after a 0.8% price move.
4. Setup filter. Confidence must exceed the threshold (default 70) and an independent ten-vote setup score (SMA 20 vs 50, RSI, MACD, Stoch RSI, volume, last-five candle colour, 20-bar support and resistance, Fear & Greed, CVD, Volume Bias) must land on BUY or SELL in the same direction. Any disagreement means no fill.

Sizing, limits and exits

Leverage is fixed at 10x virtual. Margin is 1% to 5% of the balance, scaled by confidence, capped at 2,000 USDT. At most five positions at once, one per pair, three in the same direction; the desk stops opening for the rest of the UTC day when the day's loss passes 1,000 USDT or equity falls 5% from the day's high; both limits start fresh the next day. Exits are ATR-based since 9 Sep 2026. The stop sits 1.5x the 1h ATR(14) away, clamped to between 1% and 3% of price, and that distance is one R. Targets are 1R (closes 25%), 2R (closes 25%) and 3.5R (closes the rest), so half the position rides the move. After the first target the stop moves to entry and a trailing stop of 1.2x ATR (0.6% to 2.5%) follows the move, tightening after the second target. Before that date the ladder was fixed: stop 2%, targets 1.5% (closing 40%), 3% (closing 30%) and 5% - a shape that peeled winners too early and is why the record below shows a high hit rate with a negative return. A server worker prices every open position from Binance every 20 seconds and applies these rules whether or not a browser is open.

What the numbers mean, honestly

Per exit, the average win is 41.17 USDT against an average loss of 95.50 USDT, a 0.43:1 ratio, so winners are being cut short relative to losers. Profit factor 0.77 (below 1 means the engine has lost money over this sample). Counting one result per closed position instead of per exit, 170 positions closed with a 43.5% win rate; the 64% headline counts every partial exit separately. Engine-only return since 2026-08-16: -21.96%. A short sample can and will regress, and a high hit rate with a flat or negative P&L means the exits are giving back what the entries earn. The decision layer only runs while the desk page is open, so the reference account has quiet days. Fills are on paper: no slippage, no funding, no fees. These are the reasons the page says "study partner", never "signal service".

About the AI tag: a fill counts as the engine's when the decision layer running in the desk owner's browser placed it while auto trading was switched on. The owner controls that switch, so the public record is the owner's own auto desk, not an unattended black box. Known limits we are working on: moving the decision layer to the server so it runs 24/7, and de-duplicating Fear & Greed, which currently counts in both the technical score and the setup filter.