ScalpingResearch
loading…

Scalping Research Program

Value-area structure × orderflow confirmation × directional bias — a hypothesis-driven search for a systematic scalping edge on Bybit perpetual futures.

Method

The trading model

  • Breakout (expansion out of value): needs orderflow confirmation — CVD expanding + volume z-score + acceptance.
  • Break-in (failed expansion → revert to value): needs exhaustion — CVD divergence / absorption, price poked the level and failed.
  • Horizon rule (from literature): minutes = reversal, 30min–1day = momentum.
  • Every result is baseline-relative (random-entry) and ATR-normalized.

Discipline

  • Uniform pass bar: n ≥ 50, edge ≥ +0.15 ATR, win@2R ≥ 30%, stable across two 30-day sub-samples.
  • Hard gate before deployment: t > 3 + walk-forward out-of-sample (Hou-Xue-Zhang / McLean-Pontiff meta-evidence).
  • Costs always netted: taker fees + slippage (~0.3%); stops 1–2% ATR-scaled.
  • Signal-only — no auto-trading.

Where it lives

J1 — Sweep-Reclaim at Prior-Day High/Low

CANDIDATE — the strongest result so far. Edge is real across the wide universe; the screener concentrates it. Cost-viable at ATR ≥ 1%.

ATR stratification

Reading it

  • Gross edge (ATR units × median ATR) is only positive-after-costs at ATR ≥ 1%.
  • At the median micro-cap ATR (~0.6%), costs eat the edge — trade only high-ATR names.
  • The high-ATR bucket ≈ your screened universe.

Screened vs unscreened

Signal vs screen

  • The edge is not a screen artifact — unscreened symbols still revert (−0.256 ATR).
  • The screen concentrates it: screened subset −0.402 vs unscreened −0.256.
  • Screened-first remains the right deployment lens.

Top symbols (event counts)

Takeaways

    L5 — Crash-Warning Model

    SURVIVOR. Predicts ≥5% / 8h drawdowns with OOS AUC 0.84+, beating the SEF-2026 reference (0.677). Passes symbol-held-out — no leakage.

    OOS AUC across evaluation modes

    Takeaways

      Practical use — risk gate

      • Cut size / flatten longs when predicted crash probability is in the top decile.
      • Top-decile ≈ 4.2–5.0× the base crash rate (7.5%).
      • Realized vol dominates — the model is essentially "high-vol regime ⇒ stand back".
      • Complement to the funding-crowding filter (L2).

      Full Strategy-Menu Analysis

      The reversion family is the consistent winner. VWAP-deviation fade (+0.65 ATR), IVA break-in fade (+0.41), CVD-divergence fade (+0.12) all positive. Breakout/momentum family is weak — except confirmed IVA breakout (+0.26).

      Ranking (mean 1h forward return, ATR units)

      Detail

      Nulls & context

      Gauges

        Hourly seasonality

        VWAP-deviation fade — cost validation

        Deviation buckets (net, taker standard)

        MAE/MFE — R-multiple & win-rate profile

        Best stop / target combo (2h path, net of cost)

        Walk-forward OOS validation (t>3 gate)

        VWAP-deviation fade

        IVA break-in fade

        Takeaways

          P0 — First-Pass Hypothesis Scorecard

          159 screened Bybit perps, 60 days of 5m bars, UTC days. Random-entry baseline mean 1h forward return −0.033 ATR.

          Strategy Menu — 32 Strategies

          Three deep-research streams reconciled: practitioner playbooks (S1–S10), academic orderflow/microstructure (S18–S24), intraday momentum/crypto (S25–S32). Evidence tiers: T1 = peer-reviewed · T2 = practitioner + institutional backtests · T3 = anecdotal. Full per-strategy detail, rules, and sources in strategy_brief.md.

          Data Layers & Next Steps

          What powers each result, and what's needed to go further.

          Data layers

          Next steps