agentleFS
ShubhamkumartrichyShubhamkumartrichy/claude-trading-skills-india/skills/backtest-lab/SKILL.md

backtest-lab

Skill64 starsChanged 16 days agoMITView on GitHub

Backtests rule-based swing ideas on NSE/BSE daily bars with Indian frictions built in: decide on the close and fill at the next open, gap-through stops, slippage, exact per-trade charges by segment and broker (STT, stamp duty, GST, DP), whole shares or lots, circuit-band no-fills, futures-only shorts and an optional STCG/LTCG estimate. Six built-in strategies; CAGR vs buy-and-hold, drawdown, Sharpe/Sortino, profit factor, expectancy in ₹ and R; parameter sweeps with a plateau-or-spike read, walk-forward and Monte Carlo. backtest.py, stdlib only, offline.

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 0 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Every strategy also takes stop_atr (k × ATR14 from the fill), stop_pct (overrides it), target_r and max_hold (bars), where 0 turns each one …

    line 54
  2. Very specific:

    Run run → sweep → walkforward → montecarlo, and stop at the first red flag.

    line 127

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Backtest Lab

A backtest cannot prove an idea works. It can show quickly that an idea does not. Judge a rule set by how little it degrades when Indian charges, overnight gaps, unseen years and bad luck are piled on, and treat a spectacular first run as a reason for suspicion.

Commands

python3 scripts/backtest.py strategies                                  # rules and defaults
python3 scripts/backtest.py run --csv tcs.csv --strategy sma-cross --params fast=20 slow=50
python3 scripts/backtest.py run --symbol POLYCAB --strategy 52w-high --risk-pct 1 --tax --trades-csv t.csv
python3 scripts/backtest.py run --csv hdfcbank.csv --strategy donchian-breakout --regime-csv nifty.csv
python3 scripts/backtest.py run --csv nifty.csv --strategy sma-cross --segment futures --lot-size 65 \
    --allow-short --leverage 3 --size-pct 300
python3 scripts/backtest.py sweep --csv tcs.csv --strategy sma-cross --grid fast=10,20,30 slow=50,100,200
python3 scripts/backtest.py walkforward --csv tcs.csv --strategy donchian-breakout \
    --grid entry_n=20,40,55 exit_n=10,20 --is-bars 756 --oos-bars 252
python3 scripts/backtest.py montecarlo --csv tcs.csv --strategy ema-pullback --sims 2000 --seed 7
python3 scripts/backtest.py montecarlo --from-trades t.csv --ruin-dd 25

Inputs follow technical-analysis: --csv, --candles (broker-MCP JSON, null candles skipped) or --symbol (yfinance, --period 10y). --start sets the first trading day while earlier bars still warm up the indicators. --grid also takes ranges (fast=10:50:10). Every command takes --json. Bad input exits 2 with one error: line.

Strategies (--strategy, override with --params key=value)

NameRule (long side; shorts mirror it)Defaults
sma-crossfast SMA crosses above slow; out when back belowfast 20, slow 50, stop 3×ATR
ema-pullbackclose above a rising trend EMA, low tags the EMA, close back above itema 20, trend 100, 2×ATR, target 2R, 20 bars
rsi-reversionRSI(period) below buy_below while above the trend SMA; out above sell_above2, 10, 70, trend 200, 10 bars
donchian-breakoutclose above the prior entry_n high; out below the prior exit_n low20 / 10, 2×ATR
bollinger-reversionclose below the lower band; out at the middle band20, k 2, trend 0 (off), 20 bars
52w-highclose above the prior lookback-bar high; out below the exit SMA252, exit_ma 50, 2.5×ATR

Every strategy also takes stop_atr (k × ATR14 from the fill), stop_pct (overrides it), target_r and max_hold (bars), where 0 turns each one off. There is no custom-code hook: new rules are added to the script as reviewed code.

How orders and costs are simulated

  • No look-ahead. Signals use bars up to and including the signal bar's close. The order fills at the next bar's open, never at the close that produced it.
  • Stops and targets rest in the market. A gap through the level fills at the open (a stop at ₹101 after a ₹96 open books ₹96). If one bar touches both, the stop is assumed to have filled first.
  • Slippage --slippage-bps (default 5) on every fill. Use 15–30 for small caps.
  • Charges come from trading-costs (charges.py) for each round trip, by --segment delivery|intraday|futures and --broker. A delivery buy sold the same day is charged as intraday.
  • Sizing: --size-pct (default 100% of equity), --risk-pct (risk to the initial stop, the position-sizing method) or --fixed-inr. The size is set on the signal close, capped by the cash available at the fill, and rounded down to whole shares or whole --lot-size lots (futures, SME).
  • Intraday turns each signal into a next-session open-to-close MIS trade.
  • Shorts need --segment futures --allow-short. An Indian cash-market short has to be bought back the same session (SLB aside), so a multi-day short only exists as futures. Futures notional ≤ equity × --leverage. Rollover, basis and margin calls are not modelled.
  • Circuits: with --circuit-band 2|5|10|20, an open at the band limit gets no fill. The entry is skipped; an exit or stop waits for the next unlocked open. Each case is listed under order events.
  • Regime filter: --regime-csv nifty.csv (or --regime-symbol NIFTY) allows longs only while the benchmark closes above its --regime-ma (200) SMA on the signal day, and shorts only below it.
  • Cash earns nothing. Sharpe and Sortino subtract --rf (default 6.5% a year, about a liquid fund or T-bill; change it).
  • --tax applies indian-tax-for-traders rates for a rough post-tax view: STCG 20% (held ≤12 months), LTCG 12.5% above ₹1.25 lakh per FY, intraday and futures at --slab-pct, plus 4% cess. It is approximate: losses are not carried forward and STT is treated as deductible, which it is not.

Reading the output

  • Versus buy-and-hold (delivery, one round trip). A rule earns its keep with a better Sharpe or Calmar, or a much smaller drawdown, not just a higher return in a bull market.
  • Max drawdown and longest underwater stretch: would you have kept following the rules through it?
  • Profit factor (after charges) below 1.2 is thin. Above 2 on fewer than 30 trades is suspicious.
  • Expectancy: about +0.2R or more after costs is worth a closer look. Breakeven trades (net ₹0) count as neither wins nor losses and do not end a losing streak.
  • Charges as % of gross profit above 25% means costs, not the market, own the result.
  • Yearly table: one great year carrying the whole test is regime luck.
  • Sweep: plateau (neighbours keep ≥75% of the best score) is what you want. Slope means pick the middle of the good zone. Spike means curve-fit. The luck hurdle is the Sharpe the best of N zero-edge variants would show by chance; the best cell should clear it comfortably.
  • Walk-forward: efficiency = out-of-sample CAGR ÷ mean in-sample CAGR. ≥0.5 holds up; below 0 fails. Frequent parameter changes mean the "optimum" is noise.
  • Monte Carlo: size for the 95th-percentile drawdown, not the backtest's one path. bootstrap varies the outcome; shuffle only reorders the trades (final equity is fixed, the path is not).

Checklist: try to break the idea before it breaks your account

  1. Look-ahead: can every input be known before the order? Today's high/low or close cannot trigger an entry today. Dividend-adjusted closes are not prices anyone traded at.
  2. Survivorship: today's NIFTY 50 or a current screener list are winners by construction. Ten years on them flatters every long strategy. Add names that were dropped, delisted or merged, or test the index itself.
  3. Overfitting: each extra grid cell is another lottery ticket. Keep grids coarse, insist on a plateau, walk forward, and leave the last 1–2 years untouched until the very end.
  4. Costs: a delivery round trip at Zerodha costs about 0.24% of the position (STT 0.1% a side, stamp duty, DP charge). 100 full-size round trips a year is about 24% of capital before any edge. High-turnover ideas die here. Rerun at 2–3× slippage and on a costlier broker.
  5. Regime: does it only work in 2020–21-style rallies? Check the yearly rows, add --regime-csv, and include a bad phase (2008, 2015–16, 2018 small caps, March 2020, late 2024–early 2025).
  6. Data: Yahoo history usually adjusts splits and bonuses, but not always. Moves of 25% or more overnight are flagged; check them. Dividends are left out for both the strategy and buy-and-hold.
  7. Small samples: under 30 trades is an anecdote. Run the same rules on several stocks and look at the spread of results, not the best one.
  8. Execution: thin names lock at circuits and slip more (--circuit-band, higher slippage). Stop orders are DAY-only on exchanges, so a real system needs GTTs (see breakout-planner).

Workflow with other skills

  1. market-regime (../market-regime/scripts/regime.py): test whether a NIFTY trend filter (--regime-csv) cuts drawdown more than it cuts return before using it live.
  2. Run run → sweep → walkforward → montecarlo, and stop at the first red flag.
  3. position-sizing (../position-sizing/scripts/risk.py): choose a --risk-pct that keeps the Monte Carlo 95th-percentile drawdown survivable.
  4. trading-discipline: --trades-csv writes journal columns, so python3 ../trading-discipline/scripts/discipline.py review t.csv breaks the backtest down by weekday and setup. Compare it with your live journal later; a large gap is a warning.
  5. breakout-planner: 52w-high and donchian-breakout are rough stand-ins for its Stage-2 entries. They tell you whether breakout buying pays on a stock at all, not whether a given VCP will work.

Limits: one symbol and one position at a time (no portfolio rotation), daily or weekly bars, no options, no tick rounding.

Output

NAME: strategy (params) | segment on NSE via broker | capital | sizing | slippage
start → end | bars (years) | signal on close, fill next open | charges as of <date>
Strategy vs Buy & hold: total return, CAGR, max DD, volatility, Sharpe (rf), Sortino, Calmar
Trades N: W / L / BE | win rate | avg win / loss | PF | expectancy ₹ (R) | payoff | losing streak
Exposure | avg hold | max DD dates | longest underwater | charges (% of gross profit) | slippage
Year table | last trades | order events | red flags | data notes

Educational analysis, not investment advice. A backtest describes the past, not your next trade.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.