From candle to trade decision,
in three honest gates.
This system finds momentum breakout candles on Binance altcoins and runs each one through a rigorous 3-step validation before it earns your capital. This doc walks you through exactly what happens to one signal — and why each gate exists.
If you've never traded before, start here.
Imagine you're at a casino, but it's one where you can study every past hand before you bet. You see a certain pattern keep appearing. Sometimes the hand wins, sometimes it loses. You want to know: over hundreds of past hands that looked just like this, did betting on it make money?
That's all this system does — but with cryptocurrency price candles instead of cards.
2. For each one, it looks at every similar candle in that coin's past and simulates 72 different trade strategies — "if I had entered here, with this stop, aiming for this profit, would I have made money?"
3. If the past says yes and a trained AI model agrees and on-chain data isn't against you, it gives you a green light with exact entry, stop, and target prices.
What's a "momentum candle"?
A single price bar where a lot happened in one direction — the body (open-to-close) is much bigger than usual, and volume was much higher than average. It's the market shouting. Most of the time, after a shout, price either keeps going (trend continuation) or fakes out and reverses. This system tries to tell you which one you're looking at.
The four words you need to know
A breakout, a pullback, and three possible endings
Below is exactly what happens in every trade this system gives you. Click Play and watch. The animation loops through the three possible outcomes — WIN, LOSS, and BREAKEVEN.
Reading the picture above
- Signal candle (green) — the big breakout bar that triggers the scanner.
- Blue dot — your entry price. On a Standard zone, price has to pull back to fill it.
- Red dashed line (SL) — where you quit if price goes against you. Fixed loss.
- Yellow dashed line (TP1) — first profit target at +1R. Often you close 50% here.
- Green dashed line (TP2) — full profit target at +2.5R (or 2.0 / 3.0 depending on the method).
- Moving white line — the actual price path through the trade.
What the whole workflow looks like end-to-end
Here are three illustrative walkthroughs showing how different signals end up with different verdicts. Numbers shown are representative of real output from the system.
Thursday afternoon. A 4-hour candle on SOL closes with 87% body and 5.1× volume on an ADX reading of 34. Regime score is 88/100 (strong GREEN). The scanner surfaces it.
Step 1 · Backtest
Step 2 · ML probability
78% (RF + isotonic, 142 analogs, all 11 features inside normal range for this setup).
Step 3 · AI verdict
"Unanimous candidate with 78% calibrated ML on a strong GREEN regime. OOS PF CI [1.4, 2.3] comfortably above 1. Pulse Intel is +9 (BULLISH: SOL CEX outflows + TVL tailwind). Take the trade."
Execution & outcome
REZ looks amazing on paper. 93% body, 6.8× volume, parabolic trend. Emotionally it screams "buy now or miss it." Let's see what the system said.
Step 1 · Backtest — the trap appears
Step 2 · ML
62% — but the filter is LOOSE and only 38 analogs exist. The probability is directional only.
Step 3 · AI verdict
"Despite raw headline numbers, three red flags: (1) LOOSE filter with n=38 is below statistical reliability, (2) honest PF = 1.0 means no actual edge after NEUTRAL correction, (3) fill rate 22% indicates severe survivor bias on the Standard zone. WFO verdict INSUFFICIENT (n_oos=3). Skip."
AVAX prints a bearish daily candle. 78% body, 2.8× volume. Regime is mixed (YELLOW, 54/100). Candidates disagree. What do we do?
Step 1 · Backtest — A and B disagree
WFO: BORDERLINE (n_oos = 6, OOS PF 1.18 · CI [0.85, 1.62]). Filter RELAXED 45%.
Step 2 · ML
ML(A) = 61%, ML(B) = 56%. Both above threshold but not high conviction.
Step 3 · AI verdict
"Candidate A has the edge on newest-bucket EVw and higher ML. WFO is BORDERLINE — edge plausible but noisy. Pulse Intel is NEUTRAL (+1). Take the trade at half size and exit on TP1 if reached; don't hold for TP2 without fresh confirmation."
Execution & outcome
It helps you say "no" to bad trades
Most breakouts on altcoins look great in the moment and fail by the next candle. The problem isn't finding breakouts — it's knowing which ones have historical edge right now, in this regime, on this coin.
So the system does three things:
Six things that make this different from backtested hopium
The scanner finds a breakout candle
Every few minutes, the scanner runs _scanner_score_signal on the last 3 closed candles of every USDT pair. If a candle meets the momentum threshold (big body, high volume, directional ADX), it becomes a candidate signal and shows up in the Scanner tab with a proposed trade plan.
What you see immediately
What happens when you click Step 1
The system scans the coin's full 1000-bar history, finds all past candles that look like this one (same body/volume profile, ratcheting down if needed to get enough analogs), and simulates 72 different trade strategies on each of them.
Those 72 come from every combination of:
- 3 entry zones — Aggressive (market), Standard (38.2% retrace), Sniper (61.8% retrace)
- 2 stop-loss styles — fixed 1.5% or ATR-based
- 4 management styles — Simple / Partial / Partial-NoBE / Trailing
- 3 take-profit levels — 2.0R / 2.5R / 3.0R
3 × 2 × 4 × 3 = 72 strategies, each tested against the same set of historical analogs.
Two candidates surface
Instead of picking one "best" method, the system surfaces two:
What's working right now on this coin. Picks the best method over the most recent time-decay bucket.
What has worked across the full history, weighted toward recent trades. Captures long-run robustness.
Walk-forward test
Right after the backtest, the system runs a rolling walk-forward: 5 different in-sample/out-of-sample cuts at 50/60/70/80/90% of the data. Each window trains on the "past" half and tests on the "future" half it hasn't seen.
Why walk-forward matters FOR CURIOUS
A regular backtest looks at the whole history at once and can accidentally cheat — it's been optimized on the same data it's tested on. Walk-forward simulates trading forward in time: you can only use data from before each trade to make the decision.
The "purged" part means training samples are dropped if their outcome would have resolved inside the test window — that would be a hidden leak of future information.
"4/5 windows with PF > 1" is much stronger evidence than "aggregate PF = 1.4" because it shows the edge holds in multiple time slices, not just one lucky cut.
How the ML model is trained
The system collects every historical candle on this coin that looks like today's signal (using the same ratchet filter), and labels each one by whether it would have won using the Candidate A method. Then it trains a model on those features — and uses the model to score today's candle.
The model type adapts to how much data is available:
| Analogs | Model | Why |
|---|---|---|
| < 20 | Heuristic rule | Not enough to train anything honest |
| < 50 | Logistic Regression | Simple, hard to overfit on tiny data |
| 50–149 | Random Forest | Good middle ground |
| ≥ 150 | Gradient Boosting | Best at learning complex patterns |
What comes out
What "RELAXED 45%" means FILTER BADGE
If your current signal is a 85%-body, 4× volume candle, the system ideally wants to train on historical analogs that are similarly explosive — say, at least 70% of that body and volume.
But very explosive signals often have few exact analogs. So the system ratchets the threshold down progressively — 70% → 55% → 45% → 35% → 25% → 20% — until it finds enough training samples. Every step is labeled:
STRICT 70% Tight analogs. Trust the probability number.
RELAXED 45% Looser match. Probability is still informative but wider.
LOOSE 20% Very broad analogs. Treat probability as directional only.
The final gate sends a structured package to a reasoning model (Groq's gpt-oss-120b by default). It sees the full diagnostic picture and produces a verdict for both candidates, then picks a winner.
What the AI receives
- Full signal characteristics (body, volume, ADX, regime)
- Both candidates' backtest stats + fill rate + decay buckets
- WFO verdict with OOS PF and confidence interval
- Both ML models' calibrated probabilities
- Canonical execution prices (it cannot make up numbers)
The verdict you get
Candidate A wins on three counts: (1) ML probability is 8pp higher (71% vs 63%), (2) OOS PF confidence interval stays above 1.0 across all 5 WFO windows, and (3) regime similarity to current conditions is 0.84 avg. Candidate B's Trailing management underperforms in high-ADX regimes like this one. Fill rate is a concern (Standard zone fills 47% of the time) but the ratchet-relaxed sample set has 94 analogs — enough to be directional. Pulse Intel adds +6 (bullish TVL delta on ETH).
When the AI says SKIP COMMON REASONS
- Zones invalid — big-body candle pushed retrace entry below the structural stop-loss. No trade possible.
- Fill rate too low — Standard/Sniper zones fill <40% on trending coins. Survivor bias warning.
- OOS PF CI straddles 1.0 — edge isn't statistically distinguishable from random.
- ML probability < 55% — model doesn't see this candle as a winner.
- Regime mismatch — historical analogs came from very different regimes; signal weak.
Every signal gives you three ways in
Each zone is a different trade-off between fill rate and reward-to-risk.
How you manage the trade once filled
Your five-point checklist before placing the order
If all five check out: place the limit order at the zone price, SL at the computed stop, TP1 at +1R (if Partial or Partial-NoBE), TP2 at the method's target R multiple. Size for your risk budget — never more than 0.5–1% per trade.
Reading the honesty signals
| Badge | What it means | Your response |
|---|---|---|
| STRICT 70% | Training analogs closely match current signal | Trust probability numerically |
| RELAXED 45% | Analogs are looser matches | Treat as directional, size normal |
| LOOSE 20% | Very broad analog set | Directional only, size smaller or skip |
| PASS WFO | 8+ OOS trades, clean edge | Green light from walk-forward |
| BORDERLINE | 5–7 OOS trades | Edge plausible but noisy |
| INSUFFICIENT | <5 OOS trades | Statistical blind spot — skip or wait |
| UNANIMOUS | Candidate A = Candidate B | Strongest signal the system gives |
| NEUTRAL 12% | 12% of past trades landed ±0.30R (flat) | Expect more "flat" than wins would suggest |
A free Nansen-lite layer over your signal
The Pulse tab blends three independent data sources into one composite score from −15 (strongly bearish) to +15 (strongly bullish):
A macro modifier (BTC.D delta + stablecoin supply trend) adds ±3 on top. Final score is ±15.
| Score | Verdict | What to do |
|---|---|---|
| +10 to +15 | STRONGLY BULLISH | Confirms long entries, cautious with shorts |
| +4 to +9 | BULLISH | Tailwind for longs |
| −3 to +3 | NEUTRAL | Pulse is not a factor — trade the structure |
| −4 to −9 | BEARISH | Headwind for longs |
| −10 to −15 | STRONGLY BEARISH | Veto for longs — wait or flip short |