Command center

Active universe (S&P 500 momentum + open positions + legacy)

Leonex rotates the universe weekly: downloads all 500 S&P components, computes 60d momentum and keeps the top-30. It ALWAYS includes tickers with open positions (to manage TP/SL/timeout even si they fall out of the top) plus legacy crypto/forex assets that we keep monitoring.

—

Carrier Health — read honesto del legacy

Síntesis cruda de los 3 indicadores de salud del carril regime_adaptive_v1_lo + Meta-Labeling. Si discrepan, este panel resume qué creer y qué no. El bridge de estrategias promovidas (Strategy Lab) tiene su propia validación independiente (DSR ≥ 0.80, walk-forward, permutación) y no se ve afectado por la salud aquí.

—

Intraday Timeframe Selector (Phase A — diagnostic)

For each asset, the Sharpe of a simple momentum strategy is computed across three timeframes (daily, 1h, 4h) and the winner is chosen. Anti-flip: it only switches from daily to intraday if the improvement exceeds 0.20 Sharpe. Current phase = diagnostic only. The rollout (having regime/signals/Meta operate on the chosen timeframe) comes later with your confirmation. Generate the card with: python agents/agente_datos_intraday.py → python agents/agente_timeframe_selector.py.

—

Strategy Lab v3 — strategies scored by their trades (anti-overfitting)

Each strategy = entry trigger + regime filter + exit config (take-profit / stop-loss / timeout). 24 entry triggers in three families — momentum/trend (volume-confirmed breakouts, aligned EMA reclaim, Donchian, MACD, ATR breakout, VWAP), mean-reversion (RSI extremes, Bollinger, down-streaks, Japanese candles), and multi-indicator confluence (entries that need two or three indicators to agree at the same time) — × 5 regime filters (strong uptrend, range-bound, pullback-in-uptrend, volatility expansion, plus unconditioned) × 4 exit configs = 480 strategies per asset. Crossing each trigger with each regime is what turns a loose indicator into a strategy with economic sense — a breakout in a trend is not the same trade as a breakout in a range. Every strategy is scored on its TRADES — each entry opens a discrete trade closed by TP / SL / timeout (Triple Barrier) — not on daily bars, so a buy-and-hold fluke with 2 trades can no longer rank high. The winner is picked by the Deflated Sharpe Ratio over trade returns (sample size = number of trades), validated walk-forward out-of-sample, and a strategy needs 40+ trades for GOLD / 25+ for SILVER. Every strategy is evaluated on 1d, 4h and 1h (the TF column) — intraday timeframes produce hundreds of trades per strategy, where the statistics are strong enough for SILVER/GOLD to be reachable honestly. Generate with: python agents/agente_strategy_lab.py.

—

Top-20 Rotation — bridge activo del momento

En vez de operar para siempre las estrategias que un dia fueron GOLD, la rotacion semanal rankea TODOS los candidatos elegibles de los 3 Labs (swing + scalping + custom) y selecciona las N mejores (default 20). Solo esas operan en el bridge. Las que salen del top quedan en histórico (active=0), no se borran. Anti-flicker: una entrada con menos de 4 semanas en el top se mantiene aunque caiga del ranking — evita "intercambios cosmeticos". Genera con python agents/agente_top20_selector.py --execute.

La rotación NO está activa todavía. Sin `agente_top20_selector` ejecutado, el bridge sigue operando todas las promovidas históricas (comportamiento anterior). Para activarla, lanza el agente con --execute.

Strategy Tracker — forward paper test of promoted strategies

When you --promote a strategy from the Strategy Lab, this tracker starts a genuine out-of-sample paper test: from that moment it counts only the trades whose entry happens after tracking began — data that was NOT used to pick the strategy. The Strategy Lab's backtest looks at the past; this measures whether the edge is real on NEW data. If a strategy holds up here its edge is genuine; if it falls apart it was overfitting that slipped through. No Alpaca orders — pure paper. Generate with python agents/agente_strategy_tracker.py.

—

CPCV + PBO (Combinatorial Purged Cross-Validation — Lopez de Prado AFML ch.12)

CPCV splits trades into N groups and runs every combination of k test groups, producing C(N,k) parallel backtest paths with purge & embargo. The result is an empirical distribution of Sharpe ratios instead of a single point. PBO (Probability of Backtest Overfitting, Bailey-Borwein-Lopez de Prado-Zhu 2014) measures how often the best in-sample configuration falls in the bottom half out-of-sample: low PBO = robust, high PBO = noise.

—

Configurations tested (Meta probability thresholds)

Drift Monitor (live system health)

Compares rolling metrics of the most recent 60 trades against historical baselines. If the system degrades enough, the PAUSE_NEW_TRADES flag activates and the executor stops opening new positions (close_monitor keeps managing open ones until they exit).

—

Risk Monitor (event-driven live)

Real-time portfolio risk engine. Tracks drawdown vs HWM, daily P&L, position concentration, leverage, distance to Triple-Barrier stop-loss and market volatility spikes. Any CRITICAL event triggers the same PAUSE_NEW_TRADES flag used by the Drift Monitor — the executor stops opening new positions while the Close Monitor keeps managing open ones. Reset with agent_risk_monitor.py --reset-pause once the cause is resolved.

Active events (this run)

HRP Sizing (Hierarchical Risk Parity)

When there are 2+ operable signals, Leonex assigns weights band risk contribution following the Lopez de Prado algorithm (AFML ch.16). Less volatile and less correlated assets get more weight. No matrix inversion — robust against estimation errors. With only 1 signal it keeps the fixed 5% sizing.

—

Causal Analysis (PC algorithm — Lopez de Prado ADIA 2026)

The Peter-Clark (PC) algorithm builds a causal DAG from conditional independence tests. It identifies which features are real CAUSES of the return (vs simple spurious correlations). This is the foundation of Causal Factor Investing: re-training Meta-Labeling restricted to causal features improves out-of-sample robustness and resists regime changes.

—

Fractional Differentiation (Lopez de Prado AFML cap. 5)

Diagnostic / preparatory. For each ticker, searches the minimum fractional order d ∈ [0, 1] that makes log(price) stationary (ADF p ≤ 0.05). With d~0.4 it is typically possible to preserve >80% of the original memory while gaining stationarity — required if we ever feed raw price as a feature into the Meta. Currently the Meta uses already-stationary features (RSI, momentum, sigma, regimes), so this card is informational. Run manually with: python agents/agente_frac_diff.py.

—

POOLED Portfolio Validation

Aggregates trades from all assets into a single chronological series and applies Walk-Forward + DSR + Bootstrap-H0. This indicator measures whether the system has edge, not whether each isolated asset has it.

—

Daily signal per asset

Joint reading of signals_report.json + validation_report.json. The verdict reflects whether the adaptive strategy passed the four filters: Walk-Forward Sharpe > 0, consistency ≥ 50%, DSR ≥ 0.95 and MC p-value < 0.05.

Adaptive strategy rules

Triple Barrier: closed trades

Lopez de Prado style labeling: every signal !=0 is simulated as a real trade with take-profit (2σ), stop-loss (1σ) and timeout (5 days). Data from the file dashboard/data/trades_report.json.

Latest closed trades

Closed by TP, SL or timeout. The reason is noted in the Exit column.

Meta-Labeling: filter trades with ML

LightGBM classifier trained with Walk-Forward on the Triple Barrier trades. It learns which configuration of regime + RSI + sigma + momentum predicts winners. The final validator uses the trades with probability ≥ 0.55. Data: dashboard/data/meta_report.json.

—

Custom / PineScript strategies — translated & rigorously evaluated

PineScript runs only inside TradingView — Leonex cannot execute it. So a strategy's logic (entry / exit rules) is translated into Leonex's model and evaluated here with the same rigour as the Strategy Lab: discrete trades (TP / SL / timeout), transaction cost deducted, walk-forward out-of-sample, and a confidence score. With 1 strategy that score is the PSR (Probabilistic Sharpe Ratio — odds the edge is real given the sample); with 2+ it becomes the DSR, which also discounts multiple testing. To add a strategy, paste its PineScript to Claude — it gets translated into agents/custom_strategies.py.

Runs only this agent — no S&P 500 re-download — so it finishes in a few minutes. Use it after pasting a new PineScript, or whenever the full pipeline did not reach the Custom Lab step.

—

—

—

—

Horizon Champions — the best rod of each category

Every non-rejected strategy classified by its design horizon — how long it holds a trade (timeframe × exit timeout). SHORT, MEDIUM and LONG compete in separate categories, so a great short-term strategy isn't buried under the long-term ones. And since short- and long-horizon edges are weakly correlated, holding the champion of each is genuine diversification.

PineScript inbox — paste a strategy to evaluate it across the universe

Paste your PineScript code below and hit Save. The server stores it to agents/pending_pinescripts/ with a timestamp. Translation to Python is still manual: after saving, paste the same code in your chat with Claude — I'll translate it into agents/custom_strategies.py. After deploying, the Custom Lab (in the Strategy Lab Swing page) re-evaluates all custom strategies across every ticker × timeframe (incl. scalping 30m / 15m / 5m) and produces the top-15 ticker breakdown you'll see in the Per-asset table below.

Strategy Lab Scalping — 30m / 15m / 5m, net of transaction costs

Same engine as the swing Strategy Lab (trigger + filter + exit, scored on TRADES, Deflated Sharpe Ratio, walk-forward out-of-sample), but on scalping timeframes and with a mandatory transaction-cost model: every trade pays a round-trip cost (spread + slippage) before Sharpe / DSR / tiers are computed — at 5min that cost decides whether a strategy wins or loses, so a cost-free backtest would be pure fiction. Honest caveat: yfinance only keeps ~60 days of intraday data = one market regime; treat these as provisional — the forward tracker validates over time. Generate with python agents/agente_strategy_lab_scalping.py.

—

Sizing engine (Thorp Kelly + Carver Vol Target)

Each trade is sized via fractional Kelly using prob_win from Meta-Labeling and the TP/SL payoff ratio of the Triple Barrier (default b=2). The aggregate portfolio is then scaled to a target annual volatility (Carver). HRP redistributes between correlated assets when 2+ signals are active.

—

Daily plan

Live signals today where Meta-Labeling predicted prob_win ≥ 0.55. Current mode: read_only. Two buttons below: the first only simulates, the second sends the orders to Alpaca paper.

Promoted strategies bridge — what's live today

For each promoted strategy (the SILVERs/GOLDs you activated with Promote), the bridge recompiles the same trigger × regime filter on the latest bar of its timeframe and checks whether the entry conditions fire today. If they do, the entry is merged into the Daily plan above so the Executor can size it and send the order. If they don't, you see exactly why not per strategy. This is the only feed of promoted-strategy signals into Alpaca paper.

—

Executor actions (history)

Table of everything the agent has decided in each execution: planned, simulated, sent, rejected or skipped. Useful as an audit.

Open positions

Cada fila lleva un boton Close que cierra solo esa posicion (sell market) sin tocar las demas. Util cuando tienes varias abiertas y quieres salir de una sin esperar al Close Monitor.

Trade Journal — decision narrative

Every Executor decision (executed, vetoed, skipped, planned) appears here with a natural-language explanation. The verdict of the qualitative Judge shows as a tag: APPROVE (green), CAUTION (amber) or VETO (red). If Ollama is running, the narratives are written by the local model; otherwise a Python template with the relevant data.

Close Monitor (live Triple Barrier)

For each open position, Leonex computes its take-profit (2σ), stop-loss (1σ) and timeout (5 days) at the moment of entry. This monitor compares the current price against those limits and closes automatically when one is hit. Without it, positions stay open forever.

Recent Orders

Setup recordatorio (Alpaca paper en EasyPanel)

1. Cuenta gratuita en alpaca.markets → "Paper Trading" (no live).
2. API Keys → Generate: anota la Key ID y la Secret.
3. EasyPanel → app leonex → Environment: añade ALPACA_API_KEY y ALPACA_API_SECRET (o ALPACA_SECRET_KEY; el cliente acepta cualquiera de los dos).
4. Save y redeploya el contenedor.
5. En este dashboard, pulsa Diagnose Alpaca connection y verifica veredict OK - all green. Si falla, el JSON del diagnóstico te dice exactamente qué env var falta o qué error devuelve Alpaca.

Agents Plan

Automation (n8n cron en EasyPanel)

Leonex corre en EasyPanel. La automatización ya NO es Windows Task Scheduler — eso pertenece a la época del PC local. Hoy los disparos del pipeline los hace n8n golpeando los endpoints HTTP del servidor (leonex_server.py) según un cron centralizado.

El cron de n8n se configura desde el workflow editor. Si añades un endpoint nuevo, recuerda actualizar el workflow correspondiente en n8n.

Working commands

Exposure per Class

How Leonex works — the data flow

From raw market data to a paper trade, and the research labs that run alongside. Everything is rebuilt by the pipeline; each box below is an agent.

Green = the core daily pipeline that ends in a paper order. Purple/blue = research labs that score and combine strategies. Amber = health checks. The labs and monitors never block the pipeline — they inform your decisions.

What each page does

Glossary

The terms that show up across Leonex, in plain language.