How AI Crypto Trading Signals Work — A Technical Breakdown

"AI-powered signals" is everywhere in 2026. But what does it actually mean? Is it just a buzzword, or is there real machine learning happening under the hood?

This article explains — technically but accessibly — how a real AI signal system works, using BullRunSignals as the example.

Step 1 — The Scanner: Finding Raw Setups

Before any AI gets involved, a scanner runs every 15 minutes across all monitored exchanges (Binance, Kraken, Bybit, OKX, Hyperliquid, Huobi). It analyses 60+ trading pairs looking for technical setups that meet a minimum bar.

The scanner checks 8 indicators simultaneously:

• RSI — oversold/overbought on 1h and 4h
• MACD — bullish/bearish zero-line crosses
• EMA 20/50 — trend direction and pullbacks
• Bollinger Bands — breakouts from lower/upper band
• ATR — volatility measurement for SL sizing
• Stochastic — divergence from price action
• Volume — spike detection for confirmation
• ADX — trend strength filter

Each indicator that confirms the setup adds to a score out of 10. A setup needs to score at least 7/10 to be passed to the ML filter. This alone rejects about 85% of raw setups.

Step 2 — Building the Feature Vector

For every setup that passes the scanner, the system builds a feature vector — a structured row of data that the ML model will evaluate. This includes:

Notice something surprising: the hour and day of the week are among the top features. This makes intuitive sense — a SHORT signal at 3am UTC on a Sunday has very different win probability than the same signal at 10am on a Tuesday (London open).

Step 3 — The ML Model: AutoGluon Ensemble

The core AI is an AutoGluon ensemble — a framework developed by Amazon that automatically trains and combines multiple ML models. The ensemble used in BullRunSignals includes:

The model outputs a single number: P(WIN) — the probability that this specific trade setup will result in a win. If P(WIN) ≥ 0.55 (55%), the signal is approved. Below that, it's rejected.

Step 4 — Exchange-Specific Rules

Before the ML even runs, hard rules filter out setups based on historical performance per exchange. For example:

• Binance: SHORT only (LONG has 19% win rate historically), score ≥ 9.5, BTC bear regime only
• Bybit: SHORT only (LONG 14% win rate), score ≥ 8.0
• Kraken: LONG allowed only when BTC is in confirmed bull regime
• Hyperliquid: 6h–12h UTC only (67% win rate vs 25% outside that window), R/R ≥ 2.5

These aren't arbitrary rules — they're derived from 1,000+ real closed trades per exchange. The ML refines further, but these exchange-specific filters are the first line of defense.

Step 5 — BTC Momentum Context

Every signal evaluation also checks BTC's current momentum in real time:

• BTC 1h change
• BTC 4h change
• EMA5 direction (UP or DOWN)

This generates a market regime: BULL, BEAR, or NEUTRE. The regime affects thresholds — for example, in a BEAR regime, LONG signals require higher confidence to be approved.

Step 6 — Retraining (Daily)

The model retrains every day at 05:00 CET on fresh trade data. As new trades close and add to the database, the model learns from them. This means the AI is constantly adapting to changing market conditions — not frozen on data from 6 months ago.

Why AI Beats Rule-Based Systems

Traditional rule-based systems (like "buy when RSI < 30") fail because markets change. A rule that worked in 2023 might be systematically wrong in 2026. ML models learn from actual outcomes — they don't care about rules, they care about what has actually worked.

The proof is in the numbers: without ML filtering, the BullRunSignals scanner produces ~40% win rate (baseline). With ML filtering at 55% threshold, that jumps to 78%+. The AI is adding real, measurable value.