Every 2 minutes: triage 92 markets, run AI consensus, apply risk gates, execute. Continuously.
DeepSeek pre-screens all 92 symbols down to 8–12 candidates — cutting ~$3.50 per scan in LLM API costs.
Claude (50%), pandas-ta (40%), GPT-4 Vision (10%) vote independently. ≥65% combined confidence required to trade.
Kelly sizing, correlation guards, trailing stops, MTF exits, circuit breakers — all enforced before the order fires.
After every trade, agent weights update based on what actually worked. The system gets measurably better each week.
Three independent AI agents analyse each candidate trade from different perspectives. A trade only executes when the combined weighted score clears 65% — below that, it's silently skipped.
Five independent risk layers run in parallel — before, during, and after every trade.
Every trade requires ≥65% combined agent confidence (Claude 50%, technical 40%, vision 10%). Anything below threshold is skipped — there are no marginal trades.
Fractional Kelly scales position size to the current win rate and average win/loss ratio. A hard per-position cap enforces a ceiling regardless of Kelly output.
Same-direction position caps adapt to the current regime: 3 in neutral, 5 in strong trend. Prevents concentrated directional bets disguised as diversified trading.
ATR-based and Supertrend stops trail each open position, locking profits as the trade moves in your favour. Stop type is selected per-trade based on signal characteristics.
Real-time self-monitoring at account and symbol level. Multiple independent triggers activate without requiring manual intervention.
92 symbols enter. A handful survive. Every gate has a reason.
All 92 symbols scored in a single batch call to DeepSeek. Prunes low-probability setups before any expensive model runs. Symbols that don’t cross the minimum confidence threshold are dropped here — instantly, at near-zero cost.
~20 symbols pass · 72 droppedSurvivors are evaluated across three timeframes. The 4h trend is a hard gatekeeper — if it disagrees with the trade direction, the symbol is eliminated regardless of any other signal. 1h confirms. 15m times the entry. Cached results mean this runs at zero API cost on warm cycles.
~8 symbols pass · hard gate eliminates the restOpen interest and funding rate checked for every candidate. Extreme crowded-long or crowded-short scores (> 0.7 or < −0.7) flag the setup as a crowded trade risk. Position sizing is reduced or the setup is skipped.
Funding threshold: 50% annualised · OI change: 24h deltaHyperliquid liquidation WebSocket monitored continuously. When forced-selling exhaustion is detected — a M+ cascade that suddenly stops — the detector fires a bottom-signal for that symbol. This is a separate alpha source that can open trades independent of the main pipeline.
70% historical win rate · 5-min lookback windowDeepSeek scores live headlines from 6 sources every 3 minutes. Strongly negative macro sentiment (e.g. Fed surprise, black swan) reduces position sizing or blocks new entries. Sentiment state decays to neutral automatically so old news doesn’t poison future scans.
Reuters · BBC · CoinDesk · CoinTelegraph · CryptoPanic · TheNewsAPIThree agents independently evaluate each candidate. The Strategy Agent scores the macro setup. The Technical Confirmation Agent evaluates indicator confluence. GPT-4 Vision renders a live price chart and identifies patterns, support/resistance levels, and risk. All three must agree above threshold before the trade is considered.
Minimum consensus threshold: 0.68 confidence · all 3 agents requiredThe stacked ensemble (GradientBoosting + RandomForest + Logistic Regression) produces a final win-probability score using historical features: symbol-specific win rate, market regime, agent confidence, funding environment, volatility regime, and more. SHAP values explain the score. The meta-learner can veto a trade even if all three agents agree.
Online SGD updates after every close · concept drift detection activeKelly Criterion sizes the position. ATR-based stops are placed. Correlation guard blocks trades that would make the portfolio too directional. Daily loss limit and max drawdown circuit breakers checked. If all pass — the order is placed on Hyperliquid or Binance, with trailing stop and take-profit orders set simultaneously.
Kelly fraction capped at 25% · max drawdown: 20% · daily loss limit: 10%Every layer of the system is purpose-built for live perpetuals trading. No shortcuts.
GradientBoosting + RandomForest + Logistic Regression with online SGD. After every trade, SHAP surfaces which features drove the call. Concept drift detection fires on regime shifts. Experience replay re-weights high-value trades.
Monitors Hyperliquid’s WebSocket liquidation feed in real time. Classifies forced-selling by intensity: weak (00k) to extreme (M+). Cascade exhaustion signals local bottoms. 70% historical accuracy.
Reuters, BBC, CoinDesk, CoinTelegraph, CryptoPanic & TheNewsAPI aggregated every 3 minutes. DeepSeek classifier scores macro sentiment. State decays to neutral automatically — no stale signals bleeding into entries.
Renders live price charts programmatically and sends them to GPT-4 Vision before every entry. Pattern detection, S/R identification, and risk scoring — a visual AI layer on top of quantitative signals.
Tracks open interest and funding rates across all markets. Detects extreme crowding before the unwind. Crowded-score runs −1.0 (extreme short) to +1.0 (extreme long). 60% win rate on reversal detection.
4h trend is a hard gatekeeper — disagreement blocks the trade, period. 1h confirms. 15m times the entry. Warm cache means zero extra API calls on repeat scans.
Crypto majors, alts, memes, US tech, meme stocks, metals, energy, indices, forex, pre-IPO perps and more. Entry thresholds, leverage caps and regime filters tuned independently per class.
/portfolio, /positions, /pnl, /limits, /kelly, /drawdown, /atr, /liquidation, /margin, /mtf. Full bot control from your phone. Every event — entry, exit, circuit breaker, daily summary — pushed instantly.
Exchange keys encrypted at rest, decrypted in-memory only at execution. Every scan logged: which symbols passed, each agent’s score, and the exact reason a trade was taken or skipped.
Here’s how Vektor stacks up against the alternatives.
| Manual Trading |
Generic Bots |
● Vektor | |
|---|---|---|---|
| Trades 24/7 without fatigue Always-on executionVektor runs continuously on a dedicated VPS. No sleep, no distractions, no missed setups at 3am. Every market session covered automatically. |
✕ | ✓ | ✓ |
| Multi-agent AI consensus 3-agent vote systemThree independent AI agents — Strategy, Technical Confirmation, and GPT-4 Vision — must all agree before a trade fires. One dissent = no trade. |
✕ | ✕ | ✓ |
| Dynamic Kelly position sizing Math-driven bet sizingPosition size is calculated from your actual win rate and avg win/loss ratio using the Kelly Criterion, capped at 25% to prevent over-concentration. |
Manual | Fixed % | ✓ |
| Regime-aware strategy switching Market context detectionThe bot classifies current market conditions (strong bull, bear, choppy, neutral) and adjusts trade bias accordingly — only longs in bulls, only shorts in bears. |
✕ | ✕ | ✓ |
| Circuit breakers & drawdown limits Auto-stop protectionDaily loss limit: 10%. Max drawdown: 20%. When either trips, the bot halts new entries automatically and sends a Telegram alert. |
Manual | Basic SL | ✓ |
| Self-improving meta-learner Learns from every tradeA stacked ML ensemble (GradientBoosting + RandomForest + LR) updates after every closed trade. SHAP explains each decision. The system gets smarter over time. |
✕ | ✕ | ✓ |
| GPT-4 Vision chart recognition AI sees the chartBefore every entry, a live candlestick chart is rendered and sent to GPT-4 Vision. It identifies patterns, support/resistance, and assigns a risk level as a final visual gate. |
✕ | ✕ | ✓ |
| Full signal transparency & logs No black boxEvery scan is logged: which symbols passed triage, each agent's confidence score, and the exact reason a trade was taken or rejected. Fully auditable. |
✕ | Limited | ✓ |
| AES-256 encrypted key vault Keys never exposedYour exchange API keys are encrypted at rest with AES-256. They're decrypted in-memory only at the moment of trade execution — never logged, never stored in plain text. |
— | Varies | ✓ |
| 92 markets across 5 asset classes Broad market coverageCrypto (major, alt, meme), US tech & meme stocks, commodities (gold, oil, copper), global indices, forex majors, and pre-IPO perpetuals — all in one system. |
Varies | Crypto only | ✓ |
| Liquidation cascade detection Forced-selling alphaMonitors Hyperliquid's live liquidation feed. When a cascade of forced selling exhausts ($1M+ in minutes then stops), the bot treats it as a local bottom signal. 70% win rate. |
✕ | ✕ | ✓ |
| OI + Funding rate crowding alerts Crowd positioning radarTracks open interest and funding rates across all markets. When longs or shorts become extremely crowded (score >0.7), position sizing is reduced or the setup is skipped entirely. |
✕ | ✕ | ✓ |
| Live news sentiment scoring Macro awarenessHeadlines from Reuters, BBC, CoinDesk, CryptoPanic and TheNewsAPI are scored by DeepSeek every 3 minutes. Strongly negative macro events reduce entries or block new positions. |
✕ | ✕ | ✓ |
| Multi-timeframe hard gating (4h/1h/15m) Trend alignment filterThe 4-hour trend must agree with the trade direction — if it doesn't, the trade is blocked regardless of all other signals. 1h confirms, 15m times the entry. |
Manual | Single TF | ✓ |
| Dynamic risk tightening on drawdown Self-protecting risk engineWhen drawdown exceeds 5%, minimum confidence thresholds rise, Kelly fraction shrinks, and max open positions reduces — automatically, without any manual intervention. |
✕ | ✕ | ✓ |
| Per-symbol block / probation system Underperformer quarantineSymbols with 3 consecutive losses get a 3-day block. Chronic underperformers get 7 or 30-day blocks. A probation mode allows limited re-entry to test recovery before full reinstatement. |
✕ | ✕ | ✓ |
“I was sceptical about AI trading bots — tried a few and got burned. Vektor is different. The consensus model means it only fires when all three agents agree. I’ve had fewer trades but a much cleaner win rate since joining the beta.”
“The circuit breakers alone are worth the subscription. My old bot blew up a 15% drawdown in one session. Vektor has been running for weeks — circuit breakers fired twice, both times correctly. Saved me real money.”
“The live terminal on the dashboard is genuinely addictive. Watching the agents debate a trade in real time and then seeing it execute — or get skipped — is something I’ve never seen in any other platform. Absolutely love it.”
“Set it up in paper mode on a Saturday morning. By Sunday it had placed 6 trades, closed 4 in profit. I didn’t touch anything. The Kelly sizing is conservative in a good way — it doesn’t YOLO your account.”
“I work a full-time job. I can’t sit at charts. Vektor trades while I’m in meetings — I get a Telegram ping when something happens. That’s it. It’s genuinely hands-off in a way most bots claim but don’t deliver.”
“The transparency is what gets me. Every scan shows you exactly which symbols passed triage, what each agent scored, and why the trade was skipped or taken. Most bots are a black box. Vektor shows its work.”
For the technically curious. Every system, every layer, every number.
Three-layer stacked ensemble: GradientBoostingClassifier (100 estimators, learning rate 0.05) + RandomForestClassifier (200 estimators) + LogisticRegression as meta-learner. Online learning via SGDClassifier with partial_fit — the model updates immediately after every trade close without needing a full refit. SHAP TreeExplainer surfaces feature importances per prediction. Bayesian uncertainty quantification estimates confidence intervals. Concept drift detection uses population stability index — when distributions shift, the model flags stale weights. Experience replay buffer stores 500 high-value trades; periodic replay sessions re-emphasise difficult cases.
Connects directly to Hyperliquid’s WebSocket liquidation stream. Every liquidation event is logged with symbol, side, size, price, and timestamp. A 5-minute rolling window aggregates long vs. short liquidation value. When the total crosses a threshold and the rate of new liquidations drops sharply, a cascade-exhaustion signal fires. Intensity is classified into four tiers.
Three-layer aggregation pipeline. Layer 1: RSS feeds from Reuters, BBC, CoinDesk, CoinTelegraph, MarketWatch (free). Layer 2: CryptoPanic API with ‘important’ filter (free tier). Layer 3: TheNewsAPI (paid). All headlines pass through a keyword pre-screen, SHA-256 deduplication (persisted to disk across restarts), and DeepSeek classifier. MacroSentimentState decays to neutral after a configurable TTL with no corroborating signals — ensuring old news can’t bleed into live entries.
Programmatically renders OHLCV candlestick charts using Pillow — including EMA overlays, volume bars, and S/R level annotations. The rendered image is base64-encoded and sent to GPT-4 Vision with a structured prompt requesting: pattern classification, support/resistance identification, risk level assessment (low/medium/high), and a confidence score. The vision score is combined with quantitative agent scores in the consensus calculation.
Polls Hyperliquid’s meta endpoint for open interest and funding rate across all symbols every 5 minutes. Funding rate is annualised for comparison. A crowded-score is computed from the combination of OI change (24h delta) and funding rate deviation. Scores beyond ±0.5 flag moderate crowding; beyond ±0.7 flag extreme crowding. Extreme crowded-long reduces long entries; extreme crowded-short reduces short entries.
Fetches 15m, 1h, and 4h candles concurrently (asyncio.Semaphore(8)) for every candidate symbol. Bullish signal: Price > EMA20 > EMA50. Bearish: Price < EMA20 < EMA50. 4h is a hard gatekeeper — disagreement kills the trade. Cache TTLs: 15m=4min, 1h=14min, 4h=59min. Warm-cache cycles consume zero API calls, enabling aggressive re-scanning.
Position sizing uses Kelly Criterion capped at 25% of Kelly to prevent over-concentration. Kelly fraction is calculated from rolling win rate and avg win/loss ratio. ATR (14-period) sets stop distances — stops scale with actual volatility, not fixed percentages. The Dynamic Risk Manager auto-tightens all thresholds when drawdown exceeds 5%: minimum confidence rises, Kelly fraction shrinks, max open positions reduces. Symbol Protection V2 tracks every symbol’s performance history and applies 3-day, 7-day, or 30-day blocks on underperformers — with a probation mode that allows limited re-entry for redemption. The Liquidation Protection Engine monitors real-time distance-to-liquidation, warns at 10%, fires critical alerts at 5%, and calculates exact margin injection needed to reach a safe distance.
Start free for 14 days. No credit card. Cancel anytime.
Everything you need to know before committing.
14-day free trial. No credit card. Full dashboard from day one.
⚠ Risk Disclosure: Trading perpetual contracts involves substantial risk of loss. Leverage amplifies both gains and losses. Past performance — including paper trading results shown on this site — is not indicative of future results. Vektor is a tool, not financial advice. Never trade with capital you cannot afford to lose. You are solely responsible for your trading decisions. Learn more in our FAQ.