vectorbt-expert

VectorBT Backtesting Expert Skill

Environment

• Python with vectorbt, pandas, numpy, plotly
• Data sources: OpenAlgo (Indian markets), DuckDB (direct database), yfinance (US/Global), CCXT (Crypto), custom providers
• DuckDB support: supports both custom DuckDB and OpenAlgo Historify format
• API keys loaded from single root .env via python-dotenv + find_dotenv() — never hardcode keys
• Technical indicators: OpenAlgo ta (DEFAULT - from openalgo import ta, 100+ indicators covering trend/momentum/volatility/volume/oscillators/statistical/hybrid). Use TA-Lib only if the user explicitly asks for TA-Lib/talib. NEVER use VectorBT built-in indicators either way.
• Specialty indicators (no TA-Lib equivalent, always openalgo.ta): Supertrend, Donchian, Ichimoku, HMA, KAMA, ALMA, ZLEMA, VWMA
• Signal cleaning: openalgo.ta for exrem, crossover, crossunder, flip (always, regardless of indicator library)
• Fee model: Indian market standard (STT + statutory charges + Rs 20/order)
• Benchmark: NIFTY 50 via OpenAlgo (NSE_INDEX) by default
• Charts: Plotly with template="plotly_dark"
• Environment variables loaded from single .env at project root via find_dotenv() (walks up from script dir)
• Scripts go in backtesting/{strategy_name}/ directories (created on-demand, not pre-created)
• Never use icons/emojis in code or logger output

Critical Rules

1. Default to OpenAlgo ta (from openalgo import ta) for ALL technical indicators (EMA, SMA, RSI, MACD, BBANDS, ATR, ADX, STDDEV, MOM, and 90+ more). Only use TA-Lib if the user explicitly requests "talib"/"TA-Lib" in their prompt. NEVER use vbt.MA.run(), vbt.RSI.run(), or any VectorBT built-in indicator with either library.
2. Always use OpenAlgo ta for indicators not in TA-Lib at all: Supertrend, Donchian, Ichimoku, HMA, KAMA, ALMA, ZLEMA, VWMA - these have no TA-Lib equivalent, so they're openalgo.ta even in a TA-Lib-opt-in script.
3. Use OpenAlgo ta for signal utilities: ta.exrem(), ta.crossover(), ta.crossunder(), ta.flip(). If openalgo.ta is not importable (standalone DuckDB), use inline exrem() fallback. See duckdb-data.
4. Always clean signals with ta.exrem() after generating raw buy/sell signals. Always .fillna(False) before exrem.
5. Market-specific fees: India (indian-market-costs), US (us-market-costs), Crypto (crypto-market-costs). Auto-select based on user's market.
6. Default benchmarks: India=NIFTY via OpenAlgo, US=S&P 500 (^GSPC), Crypto=Bitcoin (BTC-USD). See data-fetching Market Selection Guide.
7. Always produce a Strategy vs Benchmark comparison table after every backtest.
8. Always explain the backtest report in plain language so even normal traders understand risk and strength.
9. Plotly candlestick charts must use xaxis type="category" to avoid weekend gaps.
10. Whole shares: Always set min_size=1, size_granularity=1 for equities.
11. DuckDB data loading: When user provides a DuckDB path, load data directly using duckdb.connect() with read_only=True. Auto-detect format: OpenAlgo Historify (table market_data, epoch timestamps) vs custom (table ohlcv, date+time columns). See duckdb-data.

Modular Rule Files

Detailed reference for each topic is in rules/:

Rule File	Topic
data-fetching	OpenAlgo (India), yfinance (US), CCXT (Crypto), custom providers, .env setup
simulation-modes	from_signals, from_orders, from_holding, direction types
position-sizing	Amount/Value/Percent/TargetPercent sizing
indicators-signals	OpenAlgo ta indicator reference (default), TA-Lib opt-in, signal generation
openalgo-ta-helpers	Complete OpenAlgo ta catalog (100+ indicators): exrem, crossover, Supertrend, Donchian, Ichimoku, MAs
stop-loss-take-profit	Fixed SL, TP, trailing stop
parameter-optimization	Broadcasting and loop-based optimization
performance-analysis	Stats, metrics, benchmark comparison, CAGR
plotting	Candlestick (category x-axis), VectorBT plots, custom Plotly
indian-market-costs	Indian market fee model by segment
us-market-costs	US market fee model (stocks, options, futures)
crypto-market-costs	Crypto fee model (spot, USDT-M, COIN-M futures)
futures-backtesting	Lot sizes (SEBI revised Dec 2025), value sizing
long-short-trading	Simultaneous long/short, direction comparison
duckdb-data	DuckDB direct loading, Historify format, auto-detect, resampling, multi-symbol
csv-data-resampling	Loading CSV, resampling with Indian market alignment
walk-forward	Walk-forward analysis, WFE ratio
robustness-testing	Monte Carlo, noise test, parameter sensitivity, delay test
pitfalls	Common mistakes and checklist before going live
strategy-catalog	Strategy reference with code snippets
openstatz-tearsheet	OpenStatz interactive offline dashboard, metrics, Monte Carlo (replaces QuantStats)

Strategy Templates (in rules/assets/)

Production-ready scripts with realistic fees, NIFTY benchmark, comparison table, and plain-language report:

Template	Path	Description
EMA Crossover	assets/ema_crossover/backtest.py	EMA 10/20 crossover
RSI	assets/rsi/backtest.py	RSI(14) oversold/overbought
Donchian	assets/donchian/backtest.py	Donchian channel breakout
Supertrend	assets/supertrend/backtest.py	Supertrend with intraday sessions
MACD	assets/macd/backtest.py	MACD signal-candle breakout
SDA2	assets/sda2/backtest.py	SDA2 trend following
Momentum	assets/momentum/backtest.py	Double momentum (MOM + MOM-of-MOM)
Dual Momentum	assets/dual_momentum/backtest.py	Quarterly ETF rotation
Buy & Hold	assets/buy_hold/backtest.py	Static multi-asset allocation
RSI Accumulation	assets/rsi_accumulation/backtest.py	Weekly RSI slab-wise accumulation
Walk-Forward	assets/walk_forward/template.py	Walk-forward analysis template
Realistic Costs	assets/realistic_costs/template.py	Transaction cost impact comparison

Quick Template: Standard Backtest Script

import os
from datetime import datetime, timedelta
from pathlib import Path

import numpy as np
import pandas as pd
import vectorbt as vbt
from dotenv import find_dotenv, load_dotenv
from openalgo import api, ta

# --- Config ---
script_dir = Path(__file__).resolve().parent
load_dotenv(find_dotenv(), override=False)

SYMBOL = "SBIN"
EXCHANGE = "NSE"
INTERVAL = "D"
INIT_CASH = 1_000_000
FEES = 0.00111              # Indian delivery equity (STT + statutory)
FIXED_FEES = 20             # Rs 20 per order
ALLOCATION = 0.75
BENCHMARK_SYMBOL = "NIFTY"
BENCHMARK_EXCHANGE = "NSE_INDEX"

# --- Fetch Data ---
client = api(
    api_key=os.getenv("OPENALGO_API_KEY"),
    host=os.getenv("OPENALGO_HOST", "http://127.0.0.1:5000"),
)

end_date = datetime.now().date()
start_date = end_date - timedelta(days=365 * 3)

df = client.history(
    symbol=SYMBOL, exchange=EXCHANGE, interval=INTERVAL,
    start_date=start_date.strftime("%Y-%m-%d"),
    end_date=end_date.strftime("%Y-%m-%d"),
)
if "timestamp" in df.columns:
    df["timestamp"] = pd.to_datetime(df["timestamp"])
    df = df.set_index("timestamp")
else:
    df.index = pd.to_datetime(df.index)
df = df.sort_index()
if df.index.tz is not None:
    df.index = df.index.tz_convert(None)

close = df["close"]

# --- Strategy: EMA Crossover (OpenAlgo ta - default indicator library) ---
ema_fast = ta.ema(close, 10)
ema_slow = ta.ema(close, 20)

buy_raw = (ema_fast > ema_slow) & (ema_fast.shift(1) <= ema_slow.shift(1))
sell_raw = (ema_fast < ema_slow) & (ema_fast.shift(1) >= ema_slow.shift(1))

entries = ta.exrem(buy_raw.fillna(False), sell_raw.fillna(False))
exits = ta.exrem(sell_raw.fillna(False), buy_raw.fillna(False))

# --- Backtest ---
pf = vbt.Portfolio.from_signals(
    close, entries, exits,
    init_cash=INIT_CASH, size=ALLOCATION, size_type="percent",
    fees=FEES, fixed_fees=FIXED_FEES, direction="longonly",
    min_size=1, size_granularity=1, freq="1D",
)

# --- Benchmark ---
df_bench = client.history(
    symbol=BENCHMARK_SYMBOL, exchange=BENCHMARK_EXCHANGE, interval=INTERVAL,
    start_date=start_date.strftime("%Y-%m-%d"),
    end_date=end_date.strftime("%Y-%m-%d"),
)
if "timestamp" in df_bench.columns:
    df_bench["timestamp"] = pd.to_datetime(df_bench["timestamp"])
    df_bench = df_bench.set_index("timestamp")
else:
    df_bench.index = pd.to_datetime(df_bench.index)
df_bench = df_bench.sort_index()
if df_bench.index.tz is not None:
    df_bench.index = df_bench.index.tz_convert(None)
bench_close = df_bench["close"].reindex(close.index).ffill().bfill()
pf_bench = vbt.Portfolio.from_holding(bench_close, init_cash=INIT_CASH, fees=FEES, freq="1D")

# --- Results ---
print(pf.stats())

# --- Strategy vs Benchmark ---
comparison = pd.DataFrame({
    "Strategy": [
        f"{pf.total_return() * 100:.2f}%", f"{pf.sharpe_ratio():.2f}",
        f"{pf.sortino_ratio():.2f}", f"{pf.max_drawdown() * 100:.2f}%",
        f"{pf.trades.win_rate() * 100:.1f}%", f"{pf.trades.count()}",
        f"{pf.trades.profit_factor():.2f}",
    ],
    f"Benchmark ({BENCHMARK_SYMBOL})": [
        f"{pf_bench.total_return() * 100:.2f}%", f"{pf_bench.sharpe_ratio():.2f}",
        f"{pf_bench.sortino_ratio():.2f}", f"{pf_bench.max_drawdown() * 100:.2f}%",
        "-", "-", "-",
    ],
}, index=["Total Return", "Sharpe Ratio", "Sortino Ratio", "Max Drawdown",
          "Win Rate", "Total Trades", "Profit Factor"])
print(comparison.to_string())

# --- Explain ---
print(f"* Total Return: {pf.total_return() * 100:.2f}% vs NIFTY {pf_bench.total_return() * 100:.2f}%")
print(f"* Max Drawdown: {pf.max_drawdown() * 100:.2f}%")
print(f"  -> On Rs {INIT_CASH:,}, worst temporary loss = Rs {abs(pf.max_drawdown()) * INIT_CASH:,.0f}")

# --- Plot ---
fig = pf.plot(subplots=['value', 'underwater', 'cum_returns'], template="plotly_dark")
fig.show()

# --- Export ---
pf.positions.records_readable.to_csv(script_dir / f"{SYMBOL}_trades.csv", index=False)

Quick Template: DuckDB Backtest Script

import datetime as dt
from pathlib import Path

import duckdb
import numpy as np
import pandas as pd
import vectorbt as vbt

try:
    # Default: OpenAlgo ta for both indicators and signal cleaning
    from openalgo import ta
    exrem = ta.exrem
    ema = ta.ema
except ImportError:
    # Fallback ONLY when the openalgo package itself is not installed
    # (standalone DuckDB with no OpenAlgo). Use TA-Lib for indicators
    # and this inline exrem() replacement for signal cleaning.
    import talib as tl

    def ema(data, period):
        return pd.Series(tl.EMA(data.values, timeperiod=period), index=data.index)

    def exrem(signal1, signal2):
        result = signal1.copy()
        active = False
        for i in range(len(signal1)):
            if active:
                result.iloc[i] = False
            if signal1.iloc[i] and not active:
                active = True
            if signal2.iloc[i]:
                active = False
        return result

# --- Config ---
SYMBOL = "SBIN"
DB_PATH = r"path/to/market_data.duckdb"
INIT_CASH = 1_000_000
FEES = 0.000225              # Intraday equity
FIXED_FEES = 20

# --- Load from DuckDB ---
con = duckdb.connect(DB_PATH, read_only=True)
df = con.execute("""
    SELECT date, time, open, high, low, close, volume
    FROM ohlcv WHERE symbol = ? ORDER BY date, time
""", [SYMBOL]).fetchdf()
con.close()

df["datetime"] = pd.to_datetime(df["date"].astype(str) + " " + df["time"].astype(str))
df = df.set_index("datetime").sort_index()
df = df.drop(columns=["date", "time"])

# --- Resample to 5min ---
df_5m = df.resample("5min", origin="start_day", offset="9h15min",
                     label="right", closed="right").agg({
    "open": "first", "high": "max", "low": "min", "close": "last", "volume": "sum"
}).dropna()
close = df_5m["close"]

# --- Strategy + Backtest (same as OpenAlgo template, but use the ema()/exrem() resolved above) ---

If the user explicitly asks for TA-Lib, skip the try/except above and import talib as tl directly instead - the exrem fallback is only for when openalgo itself is unavailable.