strategy-backtesting

Strategy backtesting

Run vetted strategies over historical bars with account-correct math: pip sizes (incl. JPY), risk-based position sizing from the actual stop distance, pip-scaled spread, and explicit fill semantics. Strategies are declarative YAML or built-in only — never run user/LLM Python code (blocked by the hard rules).

When to use

• You want a statistical read (win rate, expectancy, profit factor, R distribution) on a rule-based idea, not a gut feeling.
• You need to compare a strategy across parameters or timeframes.
• You have a declarative spec and want to run it over fresh data.

Tools

• tv_backtest_run — run a built-in strategy (sma_cross, breakout, smc_h4_m15) with full currency/spread/fill controls.
• tv_backtest_render_trades — same backtest, but renders the most recent N closed trades to PNGs (entry line, SL red, TP green, exit labeled with R, band over the trade's lifetime) for a visual sanity check of the fills; needs headless Chromium. extra_markup_json layers custom drawings (killzones, boxes, text, markers, hex colors — full tv_chart_render schema) onto every image.
• Fully custom screenshots: tv_backtest_run gives every trade's entry/exit times + prices + sl/tp; feed them to tv_chart_render with end_time (window any historical trade) and your own markup_json. A user's preferred style (colors, what to draw, sizes) belongs in a skill file so the agent applies it every time — that's the customization path, no code changes.
• tv_strategy_list — list declarative YAML strategies in TV_STRATEGY_DIR.
• tv_strategy_run — run a declarative YAML strategy (same engine as backtest).
• tv_data_get_bars — fetch the bars you intend to test.

Honest interpretation checklist

1. Fill model: default is next-bar-open (trade_on_close=false). Trades that signal on a bar fill at the next open — this is the conservative real-world default. trade_on_close=true fills at the signal bar close (optimistic; only if your execution genuinely works that way).
2. Look-ahead: ensure the strategy only used information available at the signal bar (the built-ins shift their indicators by one bar). Never test a rule that references the outcome bar.
3. Sample size: per setup, fewer than ~30 trades is not evidence. Report # Trades with every conclusion; a 100% win rate on 4 trades is noise.
4. Currency: account_currency and quote_to_account_rate are explicit. When the quote (e.g. JPY) differs from the account currency, supply the rate — the tool refuses to guess.
5. Spread: spread_pips is modeled as a relative rate exact at a reference price; it is applied once at entry (backtesting.py semantics). Treat reported P&L as pre-tolerance, not a broker quote.
6. R: each trade's r = pnl / (size * |entry - sl|); a stop-hit is ~-1.0R. Judge a strategy by expectancy in R, not by total P&L (account size is arbitrary).