backtest-interpretation

Backtest Interpretation

The server's historical_price_analyzer can run four pre-built strategies over any NSE/BSE instrument with intraday or daily candles. This skill helps the user read the output without falling into the classic backtesting traps (over-fitting, survivorship, look-ahead).

1. What the tool supports

Four built-in strategies (the server's code is the source of truth):

2. Run the backtest

1. Get the user's strategy choice + parameters (periods, thresholds).
2. Call historical_price_analyzer with:
   • exchange, tradingsymbol
   • strategy (one of the four above)
   • from_date, to_date
   • interval (minute, day, etc.)
   • Strategy-specific parameters (e.g., SMA fast period, RSI period)
3. It returns trade-by-trade output plus summary stats.

3. Read the output like a skeptic

## Backtest: <strategy> on <symbol>, <from> to <to>

### Summary stats
- Total trades: <n>
- Win rate: <pct>%
- Avg P&L per trade: <amt> (<pct>%)
- Avg hold: <days>
- Sharpe ratio: <value>
- Max drawdown: <pct>%
- Final equity curve: <start> → <end> (<pct>%)

### Context needed to interpret
- Benchmark: what did buy-and-hold NIFTY 50 do over the same window? If the strategy made 15%
  and NIFTY made 20%, the strategy UNDERPERFORMED even if the number looks good.
- Time window bias: a bull-market window flatters trend-following; a choppy window flatters
  mean reversion.
- Number of trades: <10 trades means the Sharpe is unreliable (small sample).
- Costs: does the backtest include brokerage, STT, slippage? If not, real-world returns are
  lower. Typical cost drag for a delivery trade: 0.1-0.3% round-trip.

### Red flags to surface
- Sharpe >3 on daily bars is suspicious; double-check for look-ahead bias
- Max drawdown <5% on a multi-year test: could be a lucky window, not a robust strategy
- Win rate >70%: check that the avg loss isn't much bigger than avg win (classic tail risk)
- All trades in one direction: strategy hasn't been stress-tested in opposite regime