🧠 How the prediction engine works

The full back-end — what data comes in, exactly what each method measures, how they combine into one score, and where it can be made smarter.

1 · Data pipeline (in → out)

MEXC REST
Min1/5/15 klines ×600→ MEXC WebSocket
live ticks · trades · candles→ Engine
7 methods per second→ Score −100…+100
+ confidence→ Your browser
ticket · chart · push

On startup each market's history is fetched natively for every timeframe (so 5m and 15m each have 600 real candles, not aggregated approximations). Then a single MEXC websocket streams live trades, ticker and the forming candle; the server recomputes all indicators once per second and a 10-minute REST re-sync self-heals any gaps. Everything you see is MEXC's own market data.

2 · The seven methods (what predicts what)

Each method looks at the candles and votes a direction with a confidence 0–100%. This is classic technical analysis — each captures one real market behaviour:

Method	What it actually measures	Weight
EMA trend	Three exponential moving averages (21/55/200). When the fast is above the slow and price is above the 200, the market is structurally rising. This is the backbone — trend is the single most reliable edge.	30
RSI 14	Relative Strength Index — is momentum healthy (55–72) or exhausted/overbought (>72). Filters out chasing a move that's already spent.	18
MACD	Difference of two EMAs vs its signal line — measures whether momentum is still accelerating. Rising histogram = fuel left in the move.	15
Bollinger	20-period band ±2σ. A tight "squeeze" that then breaks a band signals a volatility expansion — often the start of a fast move. High confidence when breaking out of a squeeze.	14
VWAP	Volume-Weighted Average Price for the session — the day's fair value. Above it, buyers are in control; below it, sellers. Institutions anchor to VWAP.	12
Volume surge	Current candle volume vs its 20-period average. A >1.8× surge confirms real conviction behind a move rather than a thin-liquidity fakeout.	10
ADX gate	Average Directional Index — trend strength, not direction. It doesn't vote; it multiplies the whole score ×1.15 in a strong trend (ADX≥25) and ×0.55 in chop (ADX<18), because signals lie in sideways markets.	gate

3 · How they become one number

The score is a weighted, confidence-scaled vote, then gated by trend strength:

for each method:  contribution = direction(−1|0|+1) × weight × (confidence / 100)
score = ( Σ contribution / Σ weight ) × 100
score = score × ADX_multiplier          // ×1.15 strong · ×0.55 chop
score = clamp(score, −100, +100)

Score	Verdict
≥ +55	STRONG LONG
+25 … +54	LONG
−24 … +24	NEUTRAL — stand aside
−54 … −25	SHORT
≤ −55	STRONG SHORT

Confidence

Separate from the score: confidence = how much the directional methods agree with the final verdict, weighted. Six of seven methods pointing the same way → high confidence. A split vote → low confidence even if the score crosses a threshold. That's the number in the "Methods used" panel.

4 · From verdict to a trade you can place

When there's a signal, the engine builds the ticket from ATR (Average True Range — the market's live volatility):

stop-loss     = entry ∓ 1.2 × ATR      // stop breathes with volatility
take-profit 1 = entry ± 1.2 × ATR      // risk:reward 1:1  (close 50%)
take-profit 2 = entry ± 2.4 × ATR      // risk:reward 1:2  (let it run)
recommended leverage = clamp( round(0.5 / stop%), 2, 10 )   // Safe mode

In Degen mode the recommended leverage scales toward the market's real max (up to 500×) by confidence — and the calculator always shows the liquidation distance versus your stop, with a red DANGER warning if leverage would liquidate you before the stop is hit.

5 · Honest limits (what it does NOT do)

It reads price action only — no order-book depth, no funding-rate positioning, no news or on-chain flow yet.
Indicators are lagging by nature — they describe what price is doing, they don't foresee shocks (liquidation cascades, listings, macro headlines).
It has no memory of its own hit-rate — every method's weight is fixed by hand, not learned from results.
It cannot predict the future. It estimates probabilities from patterns that tend to repeat. Leverage magnifies being wrong.

6 · "Price went up but it still says SHORT?!"

This is normal for a trend-following system and usually correct. The verdict reads the whole trend (200-EMA, MACD, VWAP) — not the last tick. In a downtrend, a small up-move is a pullback, and selling into a bounce is often a better short entry, not a reason to flip.

So the dashboard now shows a separate short-term momentum line on the ticket. When price is moving against the signal, you get an amber counter-move warning: the trend still favours the trade, but wait for the short-term move to stall before entering — and if it keeps pushing against the signal, treat it as a possible early reversal and stand aside. The entry price on the ticket is now the live order-book fill price (bid for shorts, ask for longs), so it matches exactly what you'd get on MEXC.

7 · How we make it smarter (the roadmap)

Multi-timeframe confluence — only fire STRONG when 5m and 15m agree; require the higher timeframe as a directional filter. Removes most chop false-signals.
Order-book & depth — MEXC streams full depth; add bid/ask imbalance and large-wall detection to catch moves the candles haven't shown yet.
Funding-rate signal — extreme funding = crowded positioning; fade it. Cheap, high-value contrarian input already in the data.
RSI / MACD divergence — price makes a new high but momentum doesn't → early reversal warning, one of the few genuinely leading signals.
Adaptive weights (self-learning) — log every signal's outcome, then re-weight each method by its recent hit-rate per market. The engine tunes itself instead of using my hand-set numbers.
Backtesting page — replay the last N days so you (and the engine) can see the win-rate, average R, and drawdown of these exact rules before trusting size.
Volatility-regime switch — use trend-following rules in trends and mean-reversion rules in ranges, chosen automatically by ADX/Bollinger width.
Machine-learning layer (later) — once enough labelled outcomes accumulate, a gradient-boosted model can combine the same features with learned, non-linear weights.

Tell me which of these to build next — backtesting and multi-timeframe confluence give the biggest accuracy jump for the least risk.