diff --git a/kennis_streamlit.py b/kennis_streamlit.py index 528e18e..160edf7 100644 --- a/kennis_streamlit.py +++ b/kennis_streamlit.py @@ -5,57 +5,58 @@ import pandas as pd from io import BytesIO import plotly.graph_objects as go -# ----- Page config and minimal styling (institutional) ----- +# ------------------ Configuración de la página ------------------ st.set_page_config(page_title="Kennis FX Markets", layout="wide", initial_sidebar_state="expanded") st.markdown( """ """, unsafe_allow_html=True, ) -# ----- Header / Hero ----- +# ------------------ Header / Hero ------------------ col1, col2 = st.columns([4,1]) with col1: st.markdown("
Kennis FX Markets
", unsafe_allow_html=True) st.markdown("
Institutional Bayesian Intelligence Engine
", unsafe_allow_html=True) - st.markdown("
Detection and strategic allocation guidance.
", unsafe_allow_html=True) + st.markdown("
Detection and strategic allocation guidance. (Edición táctica — swing / corto plazo)
", unsafe_allow_html=True) with col2: st.markdown("
Kennis FX — Tactical edition
", unsafe_allow_html=True) st.markdown("---") -# ----- Sidebar: inputs & upload ----- -st.sidebar.header("Live inputs (tactical / swing)") -uploaded = st.sidebar.file_uploader("Upload historical CSV (optional). Required columns: cot_long,cot_short,fed_prob,retail_pct,target_up", type=["csv"]) -st.sidebar.markdown("Or use the live controls below:") +# ------------------ Sidebar: Inputs & opciones ------------------ +st.sidebar.header("Entradas (táctica / swing)") +uploaded = st.sidebar.file_uploader("Sube CSV histórico (opcional). Columnas requeridas: date,cot_long,cot_short,fed_prob,retail_pct,target_up", type=["csv"]) +st.sidebar.markdown("O ingresa la fila live abajo:") +par = st.sidebar.selectbox("Par de divisas", ["USD/JPY", "EUR/USD", "GBP/USD", "AUD/USD", "USD/CHF"], index=0) cot_long = st.sidebar.number_input("COT No-Commercial — Long", value=13662, step=1) cot_short = st.sidebar.number_input("COT No-Commercial — Short", value=13546, step=1) -fed_prob = st.sidebar.number_input("FedWatch Prob (%) for current bin", min_value=0.0, max_value=100.0, value=96.0, step=0.1) +fed_prob = st.sidebar.number_input("FedWatch Prob (%) para bin actual", min_value=0.0, max_value=100.0, value=96.0, step=0.1) retail_pct = st.sidebar.number_input("Retail % Buy (0-100)", min_value=0.0, max_value=100.0, value=47.0, step=0.1) st.sidebar.markdown("---") -st.sidebar.header("Model & normalization") -s_cot = st.sidebar.number_input("S_cot (scale for COT)", value=50000, step=1000) -prior_var = st.sidebar.number_input("Prior variance (gauss)", value=0.5, step=0.1) -retrain = st.sidebar.button("Retrain model (if CSV uploaded)") +st.sidebar.header("Modelo y normalización") +s_cot = st.sidebar.number_input("S_cot (escala para COT)", value=50000, step=1000) +prior_var = st.sidebar.number_input("Varianza prior gaussiano", value=0.5, step=0.1) +retrain = st.sidebar.button("Reentrenar (si subes CSV)") -# ----- Utilities ----- +# ------------------ Utilidades matemáticas ------------------ def sigmoid(x): return 1.0 / (1.0 + np.exp(-x)) def fit_bayesian_logit(X, y, prior_var=0.5, maxiter=200, tol=1e-8): - # MAP (Newton-Raphson) with Gaussian prior N(0, prior_var) n, p = X.shape beta = np.zeros(p) prior_prec = np.eye(p) / prior_var @@ -78,18 +79,23 @@ def fit_bayesian_logit(X, y, prior_var=0.5, maxiter=200, tol=1e-8): cov = np.linalg.inv(H) return beta, cov -def to_excel(df): +def to_excel_bytes(df): out = BytesIO() + # pandas con openpyxl como engine — openpyxl debe estar en requirements.txt with pd.ExcelWriter(out, engine="openpyxl") as writer: df.to_excel(writer, index=False, sheet_name="signal") return out.getvalue() -# ----- Data ingestion and feature building ----- +# ------------------ Ingesta de datos ------------------ if uploaded is not None: - df = pd.read_csv(uploaded) - st.info(f"CSV loaded: {uploaded.name} — {len(df)} rows") + try: + df = pd.read_csv(uploaded) + st.info(f"CSV cargado: {uploaded.name} — {len(df)} filas") + except Exception as e: + st.error("Error leyendo CSV: " + str(e)) + df = pd.DataFrame() else: - # synthetic demo history so app is usable without CSV + # demo sintético (multi-par no real) — solo para UI y pruebas np.random.seed(42) N = 420 cot_long_demo = np.random.normal(12000, 3000, size=N).astype(int) @@ -102,17 +108,18 @@ else: prob_demo = sigmoid(true_score) target_demo = (np.random.rand(N) < prob_demo).astype(int) df = pd.DataFrame({ + "date": pd.date_range(end=pd.Timestamp.today(), periods=N), "cot_long": cot_long_demo, "cot_short": cot_short_demo, "fed_prob": fed_demo, "retail_pct": retail_demo, "target_up": target_demo }) - st.info("Running on demo historical data. Upload CSV for real-data calibration.") + st.info("Usando datos demo. Suba CSV histórico para calibración con datos reales.") +# ------------------ Feature engineering ------------------ def build_features(df_in, s_cot_local=50000): df = df_in.copy() df["net"] = df["cot_long"] - df["cot_short"] df["net_scaled"] = df["net"] / float(s_cot_local) - # rolling z using up to 252 rows (business-year style) df["z_net"] = (df["net"] - df["net"].rolling(252, min_periods=1).mean()) / df["net"].rolling(252, min_periods=1).std().replace(0,1) df["delta_1w"] = df["net"].diff(5).fillna(0) df["delta_4w"] = df["net"].diff(20).fillna(0) @@ -124,37 +131,38 @@ def build_features(df_in, s_cot_local=50000): df_feat = build_features(df, s_cot_local=s_cot) -# features chosen for the tactical model +# Features used in el modelo táctico feature_cols = ["z_net", "z_delta_1w", "z_delta_4w", "signal_fed", "signal_retail"] -# prepare training data (if uploaded and has target_up will be used) if "target_up" in df_feat.columns: train_df = df_feat.copy() X_train = train_df[feature_cols].fillna(0).values y_train = train_df["target_up"].astype(int).values else: - # fallback (demo includes target_up) train_df = df_feat.copy() X_train = train_df[feature_cols].fillna(0).values y_train = train_df["target_up"].astype(int).values -# If user pressed retrain and uploaded provided CSV, we would re-fit (button here for UI clarity) if retrain and uploaded is None: - st.warning("No CSV uploaded — retrain requires historical CSV with 'target_up' labels.") -# Fit model (MAP + Laplace) -with st.spinner("Calibrating Bayesian tactical model (MAP + Laplace)..."): - beta_map, cov_post = fit_bayesian_logit(X_train, y_train, prior_var=prior_var) + st.warning("Para reentrenar debe subir un CSV histórico con etiqueta 'target_up'.") -# ----- Live row features (from sidebar) ----- +with st.spinner("Calibrando modelo bayesiano táctico (MAP + Laplace)..."): + try: + beta_map, cov_post = fit_bayesian_logit(X_train, y_train, prior_var=prior_var) + except Exception as e: + st.error("Fallo en calibración del modelo: " + str(e)) + # fallback: coeficientes neutros + beta_map = np.zeros(len(feature_cols)) + cov_post = np.eye(len(feature_cols)) * 1.0 + +# ------------------ Fila en vivo y predicción ------------------ live = {"cot_long": float(cot_long), "cot_short": float(cot_short), "fed_prob": float(fed_prob), "retail_pct": float(retail_pct)} -# to compute z etc we build by appending the live row to the history head aug = pd.concat([df.head(1).copy(), pd.DataFrame([live])], ignore_index=True) live_feat = build_features(aug, s_cot_local=s_cot).iloc[-1] x_live = live_feat[feature_cols].fillna(0).values net = live_feat["net"] -# ----- Posterior sampling and predictive distribution ----- -# defensive: ensure covariance is PSD-ish +# Posterior sampling + predicción (defensiva) try: samples = np.random.multivariate_normal(beta_map, cov_post, size=4000) sample_probs = sigmoid(samples.dot(x_live)) @@ -162,17 +170,17 @@ try: p_low = float(np.percentile(sample_probs, 2.5)) p_high = float(np.percentile(sample_probs, 97.5)) except Exception: - # fallback to point probability eta = float(x_live.dot(beta_map)) p_mean = float(sigmoid(eta)) p_low, p_high = p_mean, p_mean + sample_probs = np.array([p_mean]) -# Conviction index: transforms p_mean and dispersion into 0-100 -dispersion = max(1e-6, float(np.std(sample_probs))) if 'sample_probs' in locals() else 0.0 +# Índice de convicción +dispersion = max(1e-9, float(np.std(sample_probs))) conviction = min(100.0, max(0.0, 100.0 * (abs(p_mean - 0.5) * 2.0) * (1.0 / (1.0 + 5.0 * dispersion)))) conviction = round(conviction, 1) -# Decision thresholds (tactical) +# Decisión táctica if p_mean >= 0.60: decision = "COMPRAR (LONG)" decision_flag = "buy" @@ -183,127 +191,122 @@ else: decision = "ESPERAR (NO OPERAR)" decision_flag = "wait" -# Compute contribution breakdown (feature * beta) and normalize for narrative +# Descomposición de contribuciones contrib_raw = np.array(beta_map) * np.array(x_live) -# normalize to percent contribution with sign if np.sum(np.abs(contrib_raw)) > 0: contrib_pct = 100.0 * contrib_raw / np.sum(np.abs(contrib_raw)) else: contrib_pct = np.zeros_like(contrib_raw) -# ----- Expert narrative generator (deterministic, audit-friendly) ----- -def narrative_tactical(p_mean, p_low, p_high, conviction, decision_flag, x_live, contrib_pct, feature_cols): - # Build human expert style narrative using magnitudes and signed contributions +# ------------------ Narrativa experta (en español) ------------------ +def narrative_tactical(p_mean, p_low, p_high, conviction, decision_flag, x_live, contrib_pct, feature_cols, par): lines = [] - # regime sentence - fed_s = x_live[3] # signal_fed - cot_s = x_live[0] # z_net + lines.append(f"Resumen ejecutivo (táctico) — activo: {par}") + lines.append(f"- Probabilidad bayesiana P({par} ↑) = {p_mean*100:.1f}% (IC95%: {p_low*100:.1f}%–{p_high*100:.1f}%). Convicción: {conviction}/100.") + fed_s = x_live[3] + cot_s = x_live[0] retail_s = x_live[4] - lines.append("Executive summary (tactical):") - lines.append(f"- Bayesian posterior P(USD/JPY ↑) = {p_mean*100:.1f}% (95% CI: {p_low*100:.1f}%–{p_high*100:.1f}%). Conviction: {conviction}/100.") - # interpretation if fed_s > 0.4: - lines.append("- Macro: FedWatch is strongly hawkish — this structurally supports USD strength versus JPY over short horizon.") + lines.append("- Macro: FedWatch muestra sesgo hawkish — esto favorece fuerza del USD frente al JPY en el horizonte táctico.") elif fed_s < -0.4: - lines.append("- Macro: FedWatch is dovish — downside pressure on USD is likely.") + lines.append("- Macro: FedWatch muestra sesgo dovish — presión a la baja sobre USD esperable.") else: - lines.append("- Macro: FedWatch neutral/moderately balanced — macro is not the dominant driver.") - # COT + lines.append("- Macro: FedWatch neutral/moderado — macro por sí sola no domina la señal táctica.") if abs(cot_s) < 0.15: - lines.append("- Positioning: COT non-commercials are currently near neutral (low extremeness); tail risk from crowded positioning is limited.") + lines.append("- Posicionamiento: COT de no-commercials cerca de neutral; riesgo de cola limitado por sobreposicionamiento.") elif cot_s >= 0.15: - lines.append("- Positioning: COT shows noticeable long bias among non-commercials — structural support exists but be wary of profit-taking risk.") + lines.append("- Posicionamiento: COT exhibe sesgo long entre no-commercials — soporte estructural, pero cuidado con toma de ganancias.") else: - lines.append("- Positioning: COT shows notable short bias — risk of short-squeeze exists if macro tilts hawkish.") - # Retail + lines.append("- Posicionamiento: COT exhibe sesgo short — riesgo de short-squeeze si macro gira hawkish.") if retail_s > 0.15: - lines.append("- Retail: retail crowd leaning long — often a contrarian warning for short-term reversals; exercise caution.") + lines.append("- Retail: minoristas inclinados a comprar — aviso contrarian en plazos cortos; prudencia en entradas agresivas.") elif retail_s < -0.15: - lines.append("- Retail: retail lean short — can be supportive for medium momentum if institutional flow aligns.") + lines.append("- Retail: minoristas inclinados a vender — puede reforzar momentum si flujo institucional coincide.") else: - lines.append("- Retail: retail positioning near balanced — not a dominant contrarian signal now.") - # Contributions - lines.append("- Signal decomposition (signed % contribution):") + lines.append("- Retail: posicionamiento minorista balanceado — señal contraria no relevante ahora.") + lines.append("- Descomposición de señales (contribución firmada en %):") for name, pct, val in zip(feature_cols, contrib_pct, x_live): sign = "+" if pct >= 0 else "-" - lines.append(f" • {name}: {sign}{abs(pct):.1f}% (value {val:.3f})") - # Decision rationale + lines.append(f" • {name}: {sign}{abs(pct):.1f}% (valor {val:.3f})") if decision_flag == "buy": - lines.append("- Tactical recommendation: Gradual LONG accumulation. Rationale: hawkish Fed probabilities dominate neutral institutional positioning; market structure allows tactical upside.") - lines.append("- Suggested execution (tactical): scale-in (3 tranches), initial exposure 0.5–1.5% notional, use volatility stop (1.0–1.5 × recent ATR) and target R:R ≥ 1:1.5.") + lines.append("- Recomendación táctica: Acumulación gradual LONG. Razón: P posterior y convicción respaldan sesgo alcista táctico.") + lines.append("- Ejecución sugerida: escala en 3 tramos; tamaño inicial 0.5–1% notional; stop técnico 1.0–1.5×ATR; objetivo R:R ≥ 1:1.5.") elif decision_flag == "sell": - lines.append("- Tactical recommendation: Consider SHORT exposure or reduce existing long exposure. Rationale: posterior favors downside with sufficient conviction.") - lines.append("- Suggested execution: tight initial sizing, stop at 1.0–1.5 × ATR, target adapt to event risk.") + lines.append("- Recomendación táctica: Considerar SHORT táctico o reducción de posiciones largas. Razón: posterior favorece la baja con convicción.") + lines.append("- Ejecución sugerida: tamaño reducido; stops ajustados; vigilar noticias macro.") else: - lines.append("- Tactical recommendation: No trade. Rationale: posterior is near-neutral and uncertainty is significant; wait for clearer regime signal or confirmatory flows.") - # Caveats - lines.append("- Caveats: model uses weekly-positioning (COT) which is lagged; complement with intraday order flow and option skew for execution decisions.") + lines.append("- Recomendación táctica: No operar. Razón: incertidumbre relevante; esperar confirmación de flujo o ruptura macro.") + lines.append("- Nota: COT es semanal (retardo); combinar con order flow intradía y skew de opciones para ejecución.") return "\n".join(lines) -narrative = narrative_tactical(p_mean, p_low, p_high, conviction, decision_flag, x_live, contrib_pct, feature_cols) +narrative = narrative_tactical(p_mean, p_low, p_high, conviction, decision_flag, x_live, contrib_pct, feature_cols, par) -# ----- UI: top result and narrative ----- +# ------------------ Interfaz: resultado y narrativa ------------------ left, right = st.columns([2,3]) with left: st.markdown("
", unsafe_allow_html=True) st.markdown(f"
{decision}
", unsafe_allow_html=True) - st.markdown(f"
Prob. (mean): {p_mean*100:.1f}%    95% CI: {p_low*100:.1f}%–{p_high*100:.1f}%
", unsafe_allow_html=True) - st.markdown(f"
Conviction: {conviction}/100
", unsafe_allow_html=True) + st.markdown(f"
Probabilidad (media): {p_mean*100:.1f}%    95% IC: {p_low*100:.1f}%–{p_high*100:.1f}%
", unsafe_allow_html=True) + st.markdown(f"
Convicción: {conviction}/100
", unsafe_allow_html=True) st.markdown("
", unsafe_allow_html=True) - st.markdown("### Tactical trade plan (concise)", unsafe_allow_html=True) + st.markdown("### Plan táctico (resumen)", unsafe_allow_html=True) if decision_flag == "buy": - st.markdown("- Scale-in (3 tranches). Initial tranche: 0.5–1% notional.", unsafe_allow_html=True) - st.markdown("- Stop: technical stop (1.0–1.5×ATR) or below nearest structure.", unsafe_allow_html=True) - st.markdown("- Target: prefer R:R ≥ 1:1.5; manage increments.", unsafe_allow_html=True) + st.markdown("- Escala en 3 tramos. Tramo inicial 0.5–1% notional.", unsafe_allow_html=True) + st.markdown("- Stop: 1.0–1.5×ATR o por debajo de soporte técnico.", unsafe_allow_html=True) + st.markdown("- Target: R:R ≥ 1:1.5.", unsafe_allow_html=True) elif decision_flag == "sell": - st.markdown("- Consider reducing longs or initiating small short (0.5–1%).", unsafe_allow_html=True) - st.markdown("- Tight stops; monitor macro headlines closely.", unsafe_allow_html=True) + st.markdown("- Reducir exposiciones largas o iniciar cortos pequeños (0.5–1%).", unsafe_allow_html=True) + st.markdown("- Stops estrictos; vigila comunicados macro.", unsafe_allow_html=True) else: - st.markdown("- No trade. Await clearer signal or follow confirmatory flow.", unsafe_allow_html=True) + st.markdown("- No operar. Esperar señal clara o confirmación de flujos.", unsafe_allow_html=True) st.markdown("
", unsafe_allow_html=True) with right: - # gauge fig = go.Figure(go.Indicator( - mode="gauge+number+delta", + mode="gauge+number", value=p_mean*100, number={'suffix': '%'}, domain={'x': [0,1], 'y': [0,1]}, - title={'text': "P(USD/JPY ↑)"}, + title={'text': f"P({par} ↑)"}, gauge={ 'axis': {'range': [0,100]}, - 'bar': {'color': "#1f7a8c"}, + 'bar': {'color': "#0f2742"}, 'steps': [ - {'range':[0,40], 'color':'#a62b2b'}, - {'range':[40,60], 'color':'#bfae59'}, - {'range':[60,100], 'color':'#2a8f6b'} + {'range':[0,40], 'color':'#d9534f'}, + {'range':[40,60], 'color':'#f0ad4e'}, + {'range':[60,100], 'color':'#2a9d8f'} ], } )) st.plotly_chart(fig, use_container_width=True) -st.markdown("### Contextual narrative — expert style", unsafe_allow_html=True) +st.markdown("### Narrativa contextual — experto", unsafe_allow_html=True) st.markdown(f"
{narrative.replace(chr(10), '
')}
", unsafe_allow_html=True) -# ----- Component table and coefficients ----- -st.markdown("### Components (normalized features) and MAP coefficients") +# ------------------ Tabla de componentes y coeficientes ------------------ +st.markdown("### Componentes (valores normalizados) y coeficientes MAP") comp_df = pd.DataFrame({ "Componente": feature_cols, - "Valor (normalized)": [f"{v:.4f}" for v in x_live], + "Valor (normalizado)": [f"{v:.4f}" for v in x_live], "beta_MAP": [float(b) for b in beta_map], - "Signed % contrib": [f"{float(c):.1f}%" for c in contrib_pct] + "Contribución firmada (%)": [f"{float(c):.1f}%" for c in contrib_pct] }) st.table(comp_df) -# ----- Export & footer ----- +# ------------------ Exportar resultado ------------------ out_df = pd.DataFrame([{ - "cot_long": cot_long, "cot_short": cot_short, "net": net, + "par": par, "cot_long": cot_long, "cot_short": cot_short, "net": net, "fed_prob": fed_prob, "retail_pct": retail_pct, "p_mean": p_mean, "p_2.5": p_low, "p_97.5": p_high, "conviction": conviction, "decision": decision }]) -st.download_button("Export result (Excel)", data=to_excel(out_df), file_name="kennis_usdjpy_signal.xlsx") + +try: + excel_bytes = to_excel_bytes(out_df) + st.download_button("Exportar resultado (Excel)", data=excel_bytes, file_name=f"kennis_{par.replace('/','')}_signal.xlsx") +except Exception as e: + st.error("Error exportando Excel: " + str(e)) st.markdown("---") -st.markdown("
Implementation: Tactical Bayesian logistic (MAP + Laplace). Replace demo with real historical CSV (with 'target_up' labels) for production-grade calibration. Use with execution policy and risk controls.
", unsafe_allow_html=True) +st.markdown("
Implementación: Logistic bayesiano táctico (MAP + Laplace). Suba CSV con 'target_up' para calibración real y backtests. Uso profesional: combine con price feed y motor de ejecución.
", unsafe_allow_html=True) st.markdown("© Kennis FX Markets — Institutional Bayesian Intelligence Engine", unsafe_allow_html=True)