Every popular crypto strategy we tested on majors was negative after costs.

The falsification set

Four rules that circulate constantly in trader communities, run on the four most-traded perpetuals, with a realistic cost model: 5.5bp fee + 3bp slippage per side, plus funding drag. Non-overlapping positions, one entry per signal, exit after 24 bars. Every row links to its full card: equity curve, in-sample vs out-of-sample split, bootstrap confidence interval and the exact list of limits.

Ranges differ per row: majors share 2025-06-01 → 2026-10-11 (11,929 hourly bars); GRASSUSDT uses its full public history 2024-11-08 → 2026-10-11 (16,823 bars). “Avg/trade” is a log return in basis points, after costs. Buy&hold is the same window, for scale.

The honest read

• Reversion works on small, violent markets and dies on majors. The same 24h-fade rule that is positive on a low-cap perp is −32bp/trade on BTC and −52bp on ETH. An edge that does not transfer is a market-specific quirk, not a strategy — which is exactly why you verify per market instead of copying a setup.
• “Positive average” is not the finish line. Our best card is still labelled not separated from luck: a positive mean that fails either the significance bar (t ≥ 2, permutation p < 0.05 under two different null hypotheses) or the out-of-sample split. Most tools would have printed a equity curve and taken your money.
• Costs are the story. BTC breakout is +6bp before costs and −13bp after. Half of the rules on this page are killed by 17bp of round-trip friction, not by the market.

What a card is

Why the null hypotheses are the product

Most backtest tools will happily tell you a rule is great. We built this because we were the ones being fooled: on a strategy we trade, the permutation test that had been telling us “this is luck, size it down” turned out to be a broken null — the surrogate sampler was supposed to draw 227 random entries and was quietly producing about 9, which widens the null distribution roughly four-fold and can fake p=0.30 out of p=0.01. Fixing the sampler changed the answer, and we only trusted it after rebuilding the whole kernel independently and reconciling trade counts, averages and t-stats against the live ledger. That is the service: the discipline that is tedious to do to yourself.

What we do not do. No signals, no “today's setup”, no portfolio advice, no copy-trading, no custody of funds, no performance promises. A card is a statistical statement about a rule under stated costs on stated data. It is not advice, and markets change.

One endpoint, one card

Post a rule; get a card. JSON, or the same card rendered as a self-contained HTML page you can keep or share. Cached answers are never billed again — including after we redeploy.

curl -X POST https://api.quantcard.cc/api/card \
  -H "X-API-Key: qc_live_..." -H "Content-Type: application/json" \
  -d '{
    "source": "bybit", "symbol": "BTCUSDT", "tf": "60",
    "start": "2025-06-01", "end": "2026-10-11",
    "strategy": "mr_percentile", "hold": 24,
    "params": { "k": 24, "pct": 80, "side": "both" },
    "fee": 5.5, "slip": 3, "funding": 0.3,
    "trials": 2000, "seed": 42, "lang": "en"
  }'

→ 200 { "headline": "NEGATIVE", "metrics": { "n": 195, "meanBp": -32,
        "t": -1.99, "permP": 0.803, "ddPct": 51, "ciLoBp": -63, "ciHiBp": -2, ... } }
   x-cards-used: 1   x-cards-limit: 150   x-cards-charged: 1

Endpoints: POST /api/card · POST /api/matrix (up to 12 rule/market combos) · GET /api/strategies · /auth, /me, /webhooks/paddle. Bring-your-own rules are on the roadmap; the strategy set above is what is public today.

Pricing

You are buying cards, not hope. One computed card = one unit; a repeat of the same parameters is a cache hit and costs nothing.

Billing and tax handled by a merchant of record (Paddle). Cancel any month; the API key stops at the end of the paid period, and your previously generated cards stay readable. See refund policy, terms and privacy.

Method, and what it cannot see

We would rather hand you the limits than have you discover them in a drawdown. Every card ships with this list, unchanged:

• Log returns; costs charged per side: fee + slippage + funding drag (funding applied symmetrically unless real rate history is supplied).
• Non-overlapping positions: signals arriving while a position is open are dropped, not stacked.
• Significance is per-trade, not per-day: t-stat on net trade returns, bootstrap 95% CI on the mean, and permutation p under two nulls — the stricter one governs.
• Splits: first 60% of trades in-sample, last 40% out-of-sample, plus four time blocks so one regime cannot carry the average.
• No order-book depth ⇒ slippage is assumed, not measured. No liquidation, margin, borrow or partial-fill modelling. Fills at bar close (or next open) — no intrabar fantasy prices.
• Selection bias is the biggest hole. If you tuned the parameters on the same history you are testing, no statistic here can undo that. The only cure is forward samples; the card says so on every run.

quantcard is a research and backtesting tool. Nothing on this page is investment, financial, tax or legal advice; nothing is a recommendation to trade any instrument; past results do not predict future returns. Crypto derivatives can lose more than you deposit.