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)

1. 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

4. Common traps to call out explicitly

Over-fitting

If the user tries many parameter combinations and picks the best, Sharpe is inflated. Rule of thumb: if you tested >5 variants, discount the Sharpe by at least 30%.

Look-ahead bias

Ensure the strategy uses only data available at decision time. If "close above 20-day high" uses today's close to decide today's entry, you can't actually trade that. The server's built-in strategies are coded to avoid this, but if the user asks for a custom twist, flag the risk.

Survivorship

The stock you're backtesting is by definition still listed. Strategies that would have blown up on delisted names won't show the damage here. Relevant for small-cap backtests.

Regime shift

2015-2020 is different from 2020-2023 is different from 2023-2026. A strategy that worked in one regime often fails in the next.

No costs

If cost drag isn't in the backtest, subtract ~0.3-0.5% per round trip for realistic numbers. The server's output may or may not include costs — check the response metadata.

5. Presenting to the user

Use this framing:

• State the numbers.
• Immediately pair them with the caveats above.
• If the user asks "should I trade this live?", the answer is never a direct yes. It's: run a

walk-forward test, paper-trade it for 1-2 months (paper_trading_toggle enables this), then commit real capital with small size.

6. Offer next steps

After showing a backtest:

1. "Want to compare against buy-and-hold on the same window?" → Run historical_price_analyzer

with no strategy (or compute buy-and-hold separately from get_historical_data output).

1. "Want to stress-test across multiple symbols?" → Run the same strategy on 5-10 stocks and

compare.

1. "Want to paper-trade before going live?" → Hand off to paper-trading setup via

paper_trading_toggle.

1. "Want to see this with different parameters?" → Grid-search, but remind the user that

picking the best variant over-fits.

7. Guardrails

• NEVER present a backtest as a prediction. Past Sharpe is not future Sharpe.
• NEVER recommend going live on a backtest alone. Paper-trading is cheap; use it.
• NEVER suppress red flags (low trade count, missing costs, suspicious Sharpe). The point of

this skill is to inject skepticism, not confirm the user's bias.

• If the user asks for a backtest over <6 months of data, note that the window is too short for

meaningful statistical significance, but run it if they want the directional read.

• If the strategy's drawdown exceeds the user's stated risk tolerance, say so plainly.