trading-platform-database-v2/schemas/06_ml_schema.sql

-- ============================================================================
-- OrbiQuant IA - Esquema ML (Machine Learning)
-- ============================================================================
-- Archivo: 06_ml_schema.sql
-- Descripción: Modelos ML, predicciones, features y métricas
-- Fecha: 2025-12-05
-- ============================================================================

SET search_path TO ml;

-- ============================================================================
-- TIPOS ENUMERADOS
-- ============================================================================

CREATE TYPE model_type_enum AS ENUM (
    'range_predictor',    -- Predicción de rangos (ΔHigh/ΔLow)
    'tpsl_classifier',    -- Clasificación TP vs SL
    'signal_generator',   -- Generador de señales
    'regime_classifier',  -- Clasificación de régimen de mercado
    'amd_detector',       -- Detector de fases AMD
    'volatility_model',   -- Modelo de volatilidad
    'ensemble'            -- Meta-modelo ensemble
);

CREATE TYPE model_status_enum AS ENUM (
    'training',
    'validating',
    'ready',
    'deployed',
    'deprecated',
    'failed'
);

-- ============================================================================
-- TABLA: models
-- Descripción: Registro de modelos ML
-- ============================================================================
CREATE TABLE IF NOT EXISTS models (
    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),

    -- Identificación
    name VARCHAR(100) NOT NULL,
    slug VARCHAR(100) NOT NULL,
    description TEXT,

    -- Tipo y versión
    model_type model_type_enum NOT NULL,
    version VARCHAR(20) NOT NULL,
    is_latest BOOLEAN DEFAULT FALSE,

    -- Símbolos y timeframes
    symbols TEXT[], -- NULL = todos
    timeframes trading.timeframe_enum[],

    -- Arquitectura
    algorithm VARCHAR(50), -- 'xgboost', 'gru', 'transformer', 'ensemble'
    architecture_config JSONB, -- Configuración de arquitectura

    -- Hiperparámetros
    hyperparameters JSONB,

    -- Features
    feature_columns TEXT[],
    feature_count INT,

    -- Artifact
    artifact_path TEXT, -- Path al modelo serializado
    artifact_size_mb DECIMAL(10,2),

    -- Estado
    status model_status_enum DEFAULT 'training',
    deployed_at TIMESTAMPTZ,

    -- Metadata
    training_duration_seconds INT,
    total_samples INT,
    created_by UUID REFERENCES public.users(id),

    created_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,
    updated_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,

    UNIQUE(slug, version)
);

CREATE INDEX idx_models_slug ON models(slug);
CREATE INDEX idx_models_type ON models(model_type);
CREATE INDEX idx_models_status ON models(status);
CREATE INDEX idx_models_latest ON models(is_latest) WHERE is_latest = TRUE;

-- ============================================================================
-- TABLA: training_runs
-- Descripción: Ejecuciones de entrenamiento
-- ============================================================================
CREATE TABLE IF NOT EXISTS training_runs (
    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
    model_id UUID REFERENCES models(id) ON DELETE SET NULL,

    -- Configuración
    run_name VARCHAR(100),
    config JSONB NOT NULL,

    -- Datos
    training_data_start DATE,
    training_data_end DATE,
    validation_data_start DATE,
    validation_data_end DATE,
    total_samples INT,
    training_samples INT,
    validation_samples INT,

    -- Walk-forward
    walk_forward_splits INT,
    walk_forward_config JSONB,

    -- Estado
    status VARCHAR(20) DEFAULT 'running', -- 'running', 'completed', 'failed', 'cancelled'
    started_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,
    completed_at TIMESTAMPTZ,
    error_message TEXT,

    -- Recursos
    gpu_used BOOLEAN DEFAULT FALSE,
    memory_peak_mb INT,
    duration_seconds INT,

    -- Logs
    logs_path TEXT,

    created_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP
);

CREATE INDEX idx_training_runs_model ON training_runs(model_id);
CREATE INDEX idx_training_runs_status ON training_runs(status);

-- ============================================================================
-- TABLA: model_metrics
-- Descripción: Métricas de rendimiento de modelos
-- ============================================================================
CREATE TABLE IF NOT EXISTS model_metrics (
    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
    model_id UUID NOT NULL REFERENCES models(id) ON DELETE CASCADE,
    training_run_id UUID REFERENCES training_runs(id),

    -- Tipo de métricas
    metric_set VARCHAR(50) NOT NULL, -- 'training', 'validation', 'test', 'production'
    split_index INT, -- Para walk-forward

    -- Métricas de regresión
    mae DECIMAL(10,6),
    rmse DECIMAL(10,6),
    mape DECIMAL(10,6),
    r2_score DECIMAL(10,6),

    -- Métricas de clasificación
    accuracy DECIMAL(5,4),
    precision_score DECIMAL(5,4),
    recall_score DECIMAL(5,4),
    f1_score DECIMAL(5,4),
    roc_auc DECIMAL(5,4),

    -- Métricas por clase
    confusion_matrix JSONB,
    classification_report JSONB,

    -- Métricas de trading
    win_rate DECIMAL(5,4),
    profit_factor DECIMAL(6,2),
    sharpe_ratio DECIMAL(6,2),
    sortino_ratio DECIMAL(6,2),
    max_drawdown DECIMAL(5,4),

    -- Feature importance
    feature_importance JSONB,

    -- Timestamp
    calculated_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,

    created_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP
);

CREATE INDEX idx_model_metrics_model ON model_metrics(model_id);
CREATE INDEX idx_model_metrics_set ON model_metrics(metric_set);

-- ============================================================================
-- TABLA: predictions
-- Descripción: Predicciones generadas
-- ============================================================================
CREATE TABLE IF NOT EXISTS predictions (
    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
    model_id UUID NOT NULL REFERENCES models(id),

    -- Contexto
    symbol VARCHAR(20) NOT NULL,
    timeframe trading.timeframe_enum NOT NULL,
    prediction_timestamp TIMESTAMPTZ NOT NULL,

    -- Input
    candle_timestamp TIMESTAMPTZ NOT NULL, -- Timestamp de la vela de entrada
    input_features JSONB, -- Features usadas (opcional, para debugging)

    -- Predicción de rango
    predicted_delta_high DECIMAL(20,8),
    predicted_delta_low DECIMAL(20,8),
    predicted_delta_high_1h DECIMAL(20,8),
    predicted_delta_low_1h DECIMAL(20,8),

    -- Clasificación de bins (ATR-based)
    predicted_high_bin INT,
    predicted_low_bin INT,
    bin_probabilities JSONB,

    -- TP/SL prediction
    prob_tp_first DECIMAL(5,4),
    rr_config VARCHAR(20), -- 'rr_2_1', 'rr_3_1'

    -- Confianza
    confidence_score DECIMAL(5,4),
    model_uncertainty DECIMAL(5,4),

    -- Contexto de mercado predicho
    predicted_amd_phase trading.amd_phase_enum,
    predicted_volatility trading.volatility_regime_enum,

    -- Resultado real (llenado posteriormente)
    actual_delta_high DECIMAL(20,8),
    actual_delta_low DECIMAL(20,8),
    actual_tp_sl_outcome VARCHAR(20), -- 'tp_hit', 'sl_hit', 'neither'
    outcome_recorded_at TIMESTAMPTZ,

    -- Error calculado
    error_high DECIMAL(20,8),
    error_low DECIMAL(20,8),

    created_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP
);

CREATE INDEX idx_predictions_model ON predictions(model_id);
CREATE INDEX idx_predictions_symbol ON predictions(symbol);
CREATE INDEX idx_predictions_timestamp ON predictions(prediction_timestamp DESC);
CREATE INDEX idx_predictions_candle ON predictions(candle_timestamp);

-- Particionamiento por fecha para mejor rendimiento
-- (En producción, considerar particionar por mes)

-- ============================================================================
-- TABLA: prediction_accuracy_daily
-- Descripción: Precisión de predicciones agregada por día
-- ============================================================================
CREATE TABLE IF NOT EXISTS prediction_accuracy_daily (
    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
    model_id UUID NOT NULL REFERENCES models(id) ON DELETE CASCADE,
    symbol VARCHAR(20) NOT NULL,
    date DATE NOT NULL,

    -- Conteos
    total_predictions INT DEFAULT 0,
    predictions_evaluated INT DEFAULT 0,

    -- Métricas de rango
    mae_high DECIMAL(10,6),
    mae_low DECIMAL(10,6),
    mape_high DECIMAL(10,6),
    mape_low DECIMAL(10,6),

    -- Métricas de TP/SL
    tp_sl_predictions INT DEFAULT 0,
    tp_correct INT DEFAULT 0,
    sl_correct INT DEFAULT 0,
    accuracy_tp_sl DECIMAL(5,4),

    -- Métricas de bins
    bin_accuracy DECIMAL(5,4),

    created_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,

    UNIQUE(model_id, symbol, date)
);

CREATE INDEX idx_pred_accuracy_model ON prediction_accuracy_daily(model_id);
CREATE INDEX idx_pred_accuracy_date ON prediction_accuracy_daily(date DESC);

-- ============================================================================
-- TABLA: feature_store
-- Descripción: Features pre-calculadas para inferencia rápida
-- ============================================================================
CREATE TABLE IF NOT EXISTS feature_store (
    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),

    symbol VARCHAR(20) NOT NULL,
    timeframe trading.timeframe_enum NOT NULL,
    candle_timestamp TIMESTAMPTZ NOT NULL,

    -- OHLCV base
    open DECIMAL(20,8) NOT NULL,
    high DECIMAL(20,8) NOT NULL,
    low DECIMAL(20,8) NOT NULL,
    close DECIMAL(20,8) NOT NULL,
    volume DECIMAL(20,4),

    -- Features calculadas (las 21 del modelo)
    features JSONB NOT NULL,

    -- Indicadores técnicos
    indicators JSONB,

    -- Validación
    is_valid BOOLEAN DEFAULT TRUE,
    validation_errors TEXT[],

    created_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,

    UNIQUE(symbol, timeframe, candle_timestamp)
);

CREATE INDEX idx_feature_store_symbol ON feature_store(symbol, timeframe);
CREATE INDEX idx_feature_store_timestamp ON feature_store(candle_timestamp DESC);

-- ============================================================================
-- TABLA: model_drift_alerts
-- Descripción: Alertas de drift de modelo
-- ============================================================================
CREATE TABLE IF NOT EXISTS model_drift_alerts (
    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
    model_id UUID NOT NULL REFERENCES models(id),

    -- Tipo de drift
    drift_type VARCHAR(50) NOT NULL, -- 'feature_drift', 'prediction_drift', 'performance_drift'

    -- Detalles
    metric_name VARCHAR(100),
    expected_value DECIMAL(10,6),
    actual_value DECIMAL(10,6),
    deviation_percent DECIMAL(10,4),

    -- Severidad
    severity VARCHAR(20), -- 'low', 'medium', 'high', 'critical'

    -- Estado
    status VARCHAR(20) DEFAULT 'active', -- 'active', 'acknowledged', 'resolved'
    acknowledged_by UUID REFERENCES public.users(id),
    acknowledged_at TIMESTAMPTZ,
    resolved_at TIMESTAMPTZ,

    -- Acción tomada
    action_taken TEXT,

    created_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP
);

CREATE INDEX idx_drift_alerts_model ON model_drift_alerts(model_id);
CREATE INDEX idx_drift_alerts_status ON model_drift_alerts(status);

-- ============================================================================
-- TABLA: ab_tests
-- Descripción: Tests A/B de modelos
-- ============================================================================
CREATE TABLE IF NOT EXISTS ab_tests (
    id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),

    name VARCHAR(100) NOT NULL,
    description TEXT,

    -- Modelos
    control_model_id UUID NOT NULL REFERENCES models(id),
    treatment_model_id UUID NOT NULL REFERENCES models(id),

    -- Configuración
    traffic_split DECIMAL(3,2) DEFAULT 0.50, -- % al tratamiento
    target_metric VARCHAR(50), -- Métrica principal a optimizar

    -- Estado
    status VARCHAR(20) DEFAULT 'draft', -- 'draft', 'running', 'paused', 'completed', 'cancelled'
    started_at TIMESTAMPTZ,
    ended_at TIMESTAMPTZ,

    -- Resultados
    control_samples INT DEFAULT 0,
    treatment_samples INT DEFAULT 0,
    control_metric_value DECIMAL(10,6),
    treatment_metric_value DECIMAL(10,6),
    statistical_significance DECIMAL(5,4),
    winner VARCHAR(20), -- 'control', 'treatment', 'inconclusive'

    created_by UUID REFERENCES public.users(id),
    created_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,
    updated_at TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP
);

CREATE INDEX idx_ab_tests_status ON ab_tests(status);

-- ============================================================================
-- TRIGGERS
-- ============================================================================

CREATE TRIGGER update_models_updated_at
    BEFORE UPDATE ON models
    FOR EACH ROW
    EXECUTE FUNCTION public.update_updated_at_column();

CREATE TRIGGER update_ab_tests_updated_at
    BEFORE UPDATE ON ab_tests
    FOR EACH ROW
    EXECUTE FUNCTION public.update_updated_at_column();

-- ============================================================================
-- FUNCIÓN: Marcar modelo como latest
-- ============================================================================
CREATE OR REPLACE FUNCTION set_model_as_latest()
RETURNS TRIGGER AS $$
BEGIN
    IF NEW.is_latest = TRUE THEN
        UPDATE ml.models
        SET is_latest = FALSE
        WHERE slug = NEW.slug
          AND id != NEW.id
          AND is_latest = TRUE;
    END IF;
    RETURN NEW;
END;
$$ LANGUAGE plpgsql;

CREATE TRIGGER ensure_single_latest_model
    AFTER INSERT OR UPDATE ON models
    FOR EACH ROW
    WHEN (NEW.is_latest = TRUE)
    EXECUTE FUNCTION set_model_as_latest();