from scalarlib import getIndexConstituents, getOHLCV, getClosestTradingDate, getFundamentalsData, query_llm, strategy from datetime import datetime, time import time as time_module import pandas as pd import numpy as np # ---------- Defaults required for output capture ---------- alert_status = 'OK' run_summary = '' layer1_shortlist = [] selected_portfolio = [] web_sentiment_summary = '' trade_results = [] risk_notes = [] llm_validation_debug = [] raw_llm_selected = [] # ---------- Strategy parameters confirmed by user ---------- MODEL = 'claude-opus-4.8' INDEX = 'SPY' CAPITAL_BASE = 50000.0 LOOKBACK_DAYS = 92 FUND_LOOKBACK_DAYS = 540 TOP_N = 5 SHORTLIST_N = 25 CASH_BUFFER_PCT = 0.03 MIN_TRADE_VALUE = 25.0 MAX_WEIGHT_PCT = 35.0 MIN_WEIGHT_PCT = 5.0 MAX_QUOTE_AGE_MINUTES = 30 try: now = datetime.now() # New York time in agent runtime today = now.strftime('%Y-%m-%d') today_trading_day = getClosestTradingDate(today, offset=0).strftime('%Y-%m-%d') market_time_ok = time(9, 35) <= now.time() <= time(15, 55) state = strategy.state() if today_trading_day != today: run_summary = 'Today is not a regular trading day; no trades placed.' elif not market_time_ok: run_summary = 'Outside safe market trading window 09:35-15:55 ET; no trades placed.' elif not state.get('tradable', False): alert_status = 'ALERT' run_summary = 'Strategy is not tradable; no trades placed. Block reason: ' + str(state.get('block_reason')) elif state.get('pending'): run_summary = 'Pending orders exist; skipped this run to avoid conflicting rebalance orders.' trade_results = state.get('pending', []) else: # At 10:00 AM, today's daily bar is incomplete; use previous completed trading day. analysis_end_ts = getClosestTradingDate(today, offset=-1) analysis_end = analysis_end_ts.strftime('%Y-%m-%d') analysis_start = (analysis_end_ts - pd.Timedelta(days=LOOKBACK_DAYS)).strftime('%Y-%m-%d') fund_start = (analysis_end_ts - pd.Timedelta(days=FUND_LOOKBACK_DAYS)).strftime('%Y-%m-%d') constituents = sorted(list(set(getIndexConstituents({'index': INDEX})))) raw = getOHLCV(tickers=constituents, start=analysis_start, end=analysis_end, timeframe='day') rows = [] for t in constituents: df = raw.get(t) if df is None or df.empty or len(df) < 40: continue df = df.sort_values('ts_recv').copy() close = pd.to_numeric(df['close'], errors='coerce').dropna() volume = pd.to_numeric(df['volume'], errors='coerce').reindex(close.index).fillna(0) if len(close) < 40 or close.iloc[-1] <= 0: continue ret = close.pct_change().dropna() if len(ret) < 30: continue r_1m = float(close.iloc[-1] / close.iloc[-22] - 1) if len(close) >= 22 else np.nan r_3m = float(close.iloc[-1] / close.iloc[0] - 1) ann_vol = float(ret.std() * np.sqrt(252)) if pd.notna(ret.std()) else np.nan max_dd = float((close / close.cummax() - 1).min()) avg_dollar_volume_20d = float((close * volume).tail(20).mean()) price_score = (0.55 * r_1m) + (0.35 * r_3m) - (0.07 * ann_vol) + (0.03 * max_dd) rows.append({ 'ticker': t, 'latest_close': float(close.iloc[-1]), 'return_1m': r_1m, 'return_3m': r_3m, 'ann_vol': ann_vol, 'max_drawdown': max_dd, 'avg_dollar_volume_20d': avg_dollar_volume_20d, 'price_score': price_score, 'data_points': int(len(close)) }) tech = pd.DataFrame(rows).replace([np.inf, -np.inf], np.nan).dropna() tech = tech[tech['avg_dollar_volume_20d'] > 25_000_000].copy() if tech.empty or len(tech) < SHORTLIST_N: alert_status = 'ALERT' run_summary = 'Insufficient technical candidate data; no trades placed.' else: # Fundamental layer is applied to a technical buffer for speed/reliability. tech_buffer = tech.sort_values('price_score', ascending=False).head(80)['ticker'].tolist() fund_data = {} for sid in ['marketcap', 'pe', 'revenue']: try: fund_data[sid] = getFundamentalsData(tech_buffer, fund_start, analysis_end, sid) except Exception: fund_data[sid] = {} fundamental_rows = [] for t in tech_buffer: latest_mcap = np.nan latest_pe = np.nan rev_growth = np.nan try: df = fund_data.get('marketcap', {}).get(t) if df is not None and not df.empty: latest_mcap = float(df.sort_values('Date')['value'].dropna().iloc[-1]) except Exception: pass try: df = fund_data.get('pe', {}).get(t) if df is not None and not df.empty: latest_pe = float(df.sort_values('Date')['value'].dropna().iloc[-1]) except Exception: pass try: df = fund_data.get('revenue', {}).get(t) if df is not None and not df.empty: dfr = df.sort_values('Date').dropna(subset=['value']) if len(dfr) >= 2: latest_rev = float(dfr['value'].iloc[-1]) prior_rev = float(dfr['value'].iloc[0]) if prior_rev > 0: rev_growth = latest_rev / prior_rev - 1 except Exception: pass fundamental_rows.append({'ticker': t, 'marketcap': latest_mcap, 'pe': latest_pe, 'revenue_growth_window': rev_growth}) fund = pd.DataFrame(fundamental_rows) combined = tech.merge(fund, on='ticker', how='inner') combined['mcap_score'] = combined['marketcap'].rank(pct=True) combined['rev_score'] = combined['revenue_growth_window'].clip(-0.5, 1.0).rank(pct=True) combined['pe_score'] = np.where((combined['pe'] > 0) & (combined['pe'] <= 80), 1.0 - (combined['pe'] / 80.0), 0.2) combined['pe_score'] = pd.Series(combined['pe_score']).fillna(0.5) combined['fundamental_score'] = ( 0.35 * combined['mcap_score'].fillna(0.5) + 0.45 * combined['rev_score'].fillna(0.5) + 0.20 * combined['pe_score'].fillna(0.5) ) combined['price_rank_score'] = combined['price_score'].rank(pct=True) combined['layer1_score'] = 0.70 * combined['price_rank_score'] + 0.30 * combined['fundamental_score'] combined = combined.sort_values('layer1_score', ascending=False).reset_index(drop=True) if combined.empty or len(combined) < TOP_N: alert_status = 'ALERT' run_summary = 'Insufficient combined technical/fundamental candidate data; no trades placed.' else: shortlist = combined.head(SHORTLIST_N).copy() llm_context = shortlist[['ticker','latest_close','return_1m','return_3m','ann_vol','max_drawdown','avg_dollar_volume_20d','marketcap','pe','revenue_growth_window','price_score','fundamental_score','layer1_score']].copy() for col in ['return_1m','return_3m','ann_vol','max_drawdown','revenue_growth_window']: llm_context[col] = (llm_context[col] * 100).round(2) llm_context['latest_close'] = llm_context['latest_close'].round(2) llm_context['avg_dollar_volume_20d'] = (llm_context['avg_dollar_volume_20d'] / 1_000_000).round(1) llm_context['marketcap'] = (llm_context['marketcap'] / 1_000_000_000).round(1) for col in ['pe','price_score','fundamental_score','layer1_score']: llm_context[col] = llm_context[col].round(4) layer1_shortlist = llm_context.to_dict('records') shortlist_ticker_text = ', '.join(llm_context['ticker'].astype(str).tolist()) sp500_ticker_text = ', '.join(constituents) prompt = f''' You are selecting an aggressive long-only ALPACA_PAPER S&P 500 portfolio. Layer 1 has shortlisted stocks using technical momentum/liquidity plus fundamental quality. Now perform Layer 2: web/news/sentiment/reasoning. Use web search to assess recent sentiment, catalysts, risks, and whether the move appears supported or dangerously overextended. Layer 1 shortlist, preferred unless web evidence strongly favors another S&P 500 name: {shortlist_ticker_text} Allowed universe is the current S&P 500 only: {sp500_ticker_text} Selection rules: - You may select stocks outside the Layer 1 shortlist ONLY if they are in the S&P 500 allowed universe and web/sentiment evidence strongly justifies inclusion. - If selecting a non-shortlist S&P 500 stock, explicitly explain why the web/reasoning layer overrides the Layer 1 shortlist. - Do NOT select non-S&P 500 names, ETFs, options, crypto, private companies, IPO tickers not in S&P 500, or ticker-like themes. Choose exactly {TOP_N} stocks and assign weights summing to 100.0. Constraints: - Long-only equities only; no options, no shorts. - Aggressive but not reckless. - Daily 10:00 AM ET rebalance. - Keep a small cash buffer; weights should sum to 100 before the separate cash buffer. - Each single-name weight must be between {MIN_WEIGHT_PCT:.0f}% and {MAX_WEIGHT_PCT:.0f}%. - Prefer reputable sources: company filings/releases, major financial news, exchange/SEC/company pages, and established broker/research outlets. Treat promotional blogs as low confidence. - Do not choose a stock solely because of viral/news hype if Layer 1 evidence is weak. Return concise reasons that explicitly mention both Layer 1 evidence and Layer 2 web/sentiment evidence. ''' llm_result = query_llm( prompt=prompt, response_schema={ 'selected': [{'ticker': '', 'weight_pct': 0.0, 'layer1_reason': '', 'web_sentiment_reason': '', 'risk': ''}], 'portfolio_rationale': '', 'web_sentiment_summary': '', 'key_risks': [''], 'excluded_due_to_web_risk': [{'ticker': '', 'reason': ''}] }, context=llm_context.to_dict('records'), model=MODEL, web_search=True, ) raw_llm_selected = llm_result.get('selected', []) selected = pd.DataFrame(raw_llm_selected) sp500_tickers = set(constituents) shortlist_tickers = set(shortlist['ticker'].tolist()) validation_reasons = [] if selected.empty or 'ticker' not in selected.columns or 'weight_pct' not in selected.columns: validation_reasons.append('LLM selected output empty or missing ticker/weight_pct.') selected = pd.DataFrame(columns=['ticker', 'weight_pct', 'layer1_reason', 'web_sentiment_reason', 'risk']) else: raw_df = selected.copy() raw_df['ticker_norm'] = raw_df['ticker'].astype(str).str.upper().str.strip() non_sp500 = raw_df[~raw_df['ticker_norm'].isin(sp500_tickers)]['ticker'].astype(str).tolist() non_shortlist = raw_df[raw_df['ticker_norm'].isin(sp500_tickers) & ~raw_df['ticker_norm'].isin(shortlist_tickers)]['ticker'].astype(str).tolist() if non_sp500: validation_reasons.append('LLM chose non-S&P 500 tickers that were rejected: ' + str(non_sp500)) if non_shortlist: validation_reasons.append('LLM chose valid S&P 500 names outside Layer 1 shortlist; accepted with web override: ' + str(non_shortlist)) selected['ticker'] = selected['ticker'].astype(str).str.upper().str.strip() selected['weight_pct'] = pd.to_numeric(selected['weight_pct'], errors='coerce') selected = selected[selected['ticker'].isin(sp500_tickers)].drop_duplicates('ticker') selected = selected.dropna(subset=['weight_pct']) selected = selected[selected['weight_pct'] > 0].head(TOP_N) # Repair partial valid LLM output by filling missing slots from Layer 1 shortlist, rather than discarding all LLM picks. if len(selected) < TOP_N: validation_reasons.append(f'Only {len(selected)} valid S&P 500 LLM picks remained after filtering; filling from Layer 1 shortlist.') existing = set(selected['ticker'].tolist()) if not selected.empty else set() fillers = combined[combined['ticker'].isin(shortlist_tickers) & ~combined['ticker'].isin(existing)].head(TOP_N - len(selected))[['ticker']].copy() fillers['weight_pct'] = np.nan fillers['layer1_reason'] = 'Layer 1 filler because LLM selected too few valid S&P 500 tickers.' fillers['web_sentiment_reason'] = '' fillers['risk'] = 'Filler added after LLM validation repair.' selected = pd.concat([selected, fillers], ignore_index=True) if selected.empty or len(selected) < TOP_N: validation_reasons.append('Repair failed; using equal-weight Layer 1 fallback.') selected = combined.head(TOP_N)[['ticker']].copy() selected['weight_pct'] = 100.0 / TOP_N selected['layer1_reason'] = 'Fallback equal-weight Layer 1 pick because LLM repair failed.' selected['web_sentiment_reason'] = '' selected['risk'] = 'LLM repair failed.' else: # Preserve LLM relative weights for valid picks when possible; assign missing/filler weights equally and normalize. if selected['weight_pct'].isna().any() or selected['weight_pct'].sum(skipna=True) <= 0: selected['weight_pct'] = 100.0 / TOP_N else: # Give NaN filler rows equal residual if any, otherwise normalize all selected weights. nan_mask = selected['weight_pct'].isna() if nan_mask.any(): used = min(float(selected.loc[~nan_mask, 'weight_pct'].sum()), 80.0) residual = max(100.0 - used, 5.0 * int(nan_mask.sum())) selected.loc[nan_mask, 'weight_pct'] = residual / int(nan_mask.sum()) selected['weight_pct'] = selected['weight_pct'] / selected['weight_pct'].sum() * 100.0 # Enforce single-name bounds by clipping and renormalizing. If still bad, equal-weight selected names. selected['weight_pct'] = selected['weight_pct'].clip(lower=MIN_WEIGHT_PCT, upper=MAX_WEIGHT_PCT) selected['weight_pct'] = selected['weight_pct'] / selected['weight_pct'].sum() * 100.0 if (selected['weight_pct'].max() > MAX_WEIGHT_PCT + 0.01) or (selected['weight_pct'].min() < MIN_WEIGHT_PCT - 0.01): validation_reasons.append('Weights still breached risk bounds after clipping; equal-weighting selected names.') selected['weight_pct'] = 100.0 / TOP_N if validation_reasons: alert_status = 'ALERT' llm_validation_debug = validation_reasons risk_notes.append('LLM validation repaired: ' + '; '.join(validation_reasons)) else: llm_validation_debug = ['LLM output passed validation.'] target_tickers = selected['ticker'].tolist() web_sentiment_summary = str(llm_result.get('web_sentiment_summary', '')) snap = strategy.getLatestSnapshot(target_tickers) state = strategy.state() nav = float(state.get('capital', {}).get('nav', CAPITAL_BASE) or CAPITAL_BASE) deployable_nav = nav * (1.0 - CASH_BUFFER_PCT) current_positions = {p['symbol']: float(p.get('qty', 0) or 0) for p in state.get('positions', [])} current_market_values = {p['symbol']: float(p.get('market_value', 0) or 0) for p in state.get('positions', [])} # Flatten positions no longer selected first. Alpaca orders are async, so wait/reconcile before new buys. sell_order_placed = False for sym, qty in list(current_positions.items()): if sym not in target_tickers and abs(qty) > 0: mv = abs(current_market_values.get(sym, 0.0)) if mv >= MIN_TRADE_VALUE: resp = strategy.execute(sym, 0) trade_results.append({'symbol': sym, 'target_qty': 0, 'action': 'flatten_unselected', 'response': resp}) if resp.get('status') in ['PENDING', 'FILLED']: sell_order_placed = True if sell_order_placed: # Give Alpaca paper market orders a short window to fill, then reconcile cash/positions. for _ in range(6): time_module.sleep(2) state = strategy.state() if not state.get('pending'): break if state.get('pending'): alert_status = 'ALERT' run_summary = 'Sell orders still pending after flatten step; skipped new buys to avoid insufficient-cash / partial-rebalance state.' trade_results = trade_results + state.get('pending', []) target_tickers = [] else: state = strategy.state() nav = float(state.get('capital', {}).get('nav', CAPITAL_BASE) or CAPITAL_BASE) deployable_nav = nav * (1.0 - CASH_BUFFER_PCT) current_positions = {p['symbol']: float(p.get('qty', 0) or 0) for p in state.get('positions', [])} current_market_values = {p['symbol']: float(p.get('market_value', 0) or 0) for p in state.get('positions', [])} # Target selected positions using live quotes. selected_portfolio = [] for _, row in selected.iterrows(): if not target_tickers: break sym = row['ticker'] weight = float(row['weight_pct']) / 100.0 q = snap.get(sym, {}) quote_ts = q.get('ts_recv') if quote_ts: try: quote_age_min = (now - pd.to_datetime(quote_ts).to_pydatetime()).total_seconds() / 60.0 if quote_age_min > MAX_QUOTE_AGE_MINUTES: alert_status = 'ALERT' trade_results.append({'symbol': sym, 'status': 'SKIPPED', 'reason': 'Stale live quote', 'quote_ts': str(quote_ts), 'quote_age_min': round(quote_age_min, 1)}) continue except Exception: risk_notes.append('Could not parse quote timestamp for ' + sym + ': ' + str(quote_ts)) price = q.get('last') or q.get('ask') or q.get('bid') if price is None or float(price) <= 0: alert_status = 'ALERT' trade_results.append({'symbol': sym, 'status': 'SKIPPED', 'reason': 'No valid live price'}) continue price = float(price) target_value = deployable_nav * weight target_qty = round(target_value / price, 2) current_qty = current_positions.get(sym, 0.0) est_trade_value = abs(target_qty - current_qty) * price selected_portfolio.append({ 'ticker': sym, 'weight_pct': round(float(row['weight_pct']), 2), 'target_value': round(target_value, 2), 'live_price': round(price, 4), 'target_qty': target_qty, 'layer1_reason': str(row.get('layer1_reason', '')), 'web_sentiment_reason': str(row.get('web_sentiment_reason', '')), 'risk': str(row.get('risk', '')) }) if target_qty <= 0: trade_results.append({'symbol': sym, 'status': 'SKIPPED', 'reason': 'Target quantity <= 0'}) continue if est_trade_value < MIN_TRADE_VALUE: trade_results.append({'symbol': sym, 'status': 'SKIPPED', 'reason': 'Adjustment below minimum trade value', 'estimated_trade_value': round(est_trade_value, 2)}) continue resp = strategy.execute(sym, target_qty) trade_results.append({'symbol': sym, 'target_qty': target_qty, 'estimated_trade_value': round(est_trade_value, 2), 'response': resp}) risk_notes = list(llm_result.get('key_risks', [])) + risk_notes excluded = llm_result.get('excluded_due_to_web_risk', []) if excluded: risk_notes.append('Excluded due to web risk: ' + str(excluded)) run_summary = ( f'Completed two-layer Fable 5 S&P 500 aggressive Top {TOP_N} rebalance. ' f'Layer 1 shortlist size: {SHORTLIST_N}; data window {analysis_start} to {analysis_end}. ' f'NAV used: ${nav:,.2f}; deployable after buffer: ${deployable_nav:,.2f}. ' f'Selected: {", ".join(target_tickers)}.' ) except Exception as e: alert_status = 'ALERT' run_summary = 'Agent failed before completing rebalance: ' + str(e) risk_notes = risk_notes + ['Execution exception: ' + str(e)]

Market	Qty	Avg Price	Mkt Value	Unrealized P&L	Side

Market	Target Qty	Status	Time (ET)

Event	Platform	Market	Side	Status	Filled	Fill Price	Value	Fees	Order ID	Reason	Time (ET)

Opus 4.8 Agentic Equity ETF

NAV History

Strategy

Open Positions

Pending Orders

Recent Activity