A high-performance analysis engine for testing CISD (Close Implies Subsequent Direction) patterns across NQ and ES.

This tool evaluates the "run rate" of the CISD pattern using a strict Barrier Problem approach: Does the price hit the target (CISD High/Low) before hitting the stop (opposite side) within the lookahead window?

Swing SMT confirmation is now backed by the local SMT library at /mnt/e/backup/code/Finance/Misc/SMT. The matching rule is left-only: a CISD at bar t counts when a same-direction Swing SMT was created on t, t-1, or t-2.

Planned future research ideas live in docs/research_backlog.md. The current research surface also includes standalone analyses for CISD-linked FVG creation/hold, sweep confirmation, and direction-specific swing-position buckets.

All figures use barrier logic: target hit before stop, lookahead = 2 bars.

4H is the weakest timeframe. 1H and 15min are the most consistent. Bullish bias is persistent across all timeframes (~4–10%).

A CISD that closes past the previous wick is the single most reliable filter.

Within-wick bearish setups on the Daily are particularly weak: ES bearish within-wick hits only 39.5% — worse than random.

Combining past-wick closure with 2–3 consecutive opposite candles consistently yields the highest hit rates:

Within-wick + 2c on Daily (ES bear) drops to 36.7% — the weakest observed bucket.

Requiring the close to exceed the previous candle's High/Low (not just the close) gives a +5–8% lift:

Consecutive candle count alone has weak and inconsistent predictive value. Hit rates are largely flat across 1–3 consecutive opposite candles, staying within ±3% of the baseline. The edge only emerges when combined with wick position (see §3).

(See standalone charts: CandleSize_All_Timeframes.png, SizeCross_All_Timeframes.png)

CISD candles with a body ≥ 1x ATR(14) show meaningfully higher hit rates than smaller candles. The cross-tab reveals:

• Big CISD + Small prev = strongest quadrant.
• Small CISD + Big prev = weakest quadrant.
• A small previous candle amplifies the advantage of a large CISD body.

(See Volume_All_Timeframes.png)

Volume ratio (CISD candle vs previous candle) has negligible impact on outcomes. Hit rates are stable across all volume buckets as long as volume ≥ 1x the prior candle.

(See SMT_CISD_All_Timeframes.png)

A co-occurring same-direction Swing SMT (divergence between NQ and ES swing highs/lows, lookback=20) is used as a confirmation filter. Results vary sharply by timeframe.

Key takeaways:

• Daily SMT sample sizes are very small (n=11–26) making results unreliable. The ES Bearish w/ SMT result (27.8%) is a notable outlier but based on only 18 events.
• 4H SMT has negligible impact — differences are within ±3pp and no directional consistency.
• 1H SMT is similarly flat — lifts of <2pp across all four instrument/direction combos.
• 15min shows the most consistent positive effect: +2–4pp across all combos with large sample sizes (n=1,200–1,400).

SMT confirmation adds the most value at 15min, where it provides a small but consistent edge across both instruments and both directions.

Per-Timeframe (5 core analyses):

Standalone — All Timeframes side-by-side:

Ensure data is in .parquet format in the data/ directory (nq_1m.parquet, es_1m.parquet).

Rebuild the interactive forward-returns report with:

python3 scripts/build_forward_returns.py

The page groups research into three families:

• core: SMT, size-vs-prev-body, wick position, and consecutive-candle buckets.
• fvg: CISD-linked FVG creation and hold-state buckets.
• structure: sweep confirmation plus the previous-bar and CISD-bar swing-position buckets.

python cisd_analysis.py

python cisd_analysis.py cisd_fvg fvg_hold cisd_fvg_interaction sweep sssf_swing

The current CLI surface is limited to the analysis keys listed below.

• FVG formation uses the standard 3-candle wick-to-wick rule.
  • Bullish: left high < right low.
  • Bearish: left low > right high.
  • mid0 means the CISD bar is the middle candle; mid1 means the next bar is the middle candle.
• FVG direction must match the parent CISD direction.
• FVG hold is measured over FVG_HOLD_LOOKAHEAD = 10 bars from the FVG middle candle with two failure modes:
  • close_through_near_edge
  • wick_break_far_extreme
• Sweep confirmation is direction-specific and binary:
  • bullish CISD checks for a low sweep
  • bearish CISD checks for a high sweep
  • the sweep can occur on any bar in [t-4, t]
  • prior swing points are sourced from a 20-bar swing lookback
• Swing-position buckets use a 3-candle swing rule (N=1):
  • bullish CISDs use swing lows
  • bearish CISDs use swing highs
  • prev_bar_is_swing corresponds to candle[-1]
  • cisd_bar_is_swing corresponds to candle[0]

Edit constants at the top of cisd_analysis.py:

• LOOKAHEAD: Bars to check ahead (Default: 2).
• MAX_CONSEC: Max consecutive candles for Markov (Default: 3).
• TIMEFRAMES: Resampling rules (Default: Daily, 4H, 1H, 15min).
• FVG_HOLD_LOOKAHEAD: Bars used to judge whether a linked FVG held (Default: 10).
• SWEEP_TOLERANCE: Bars in the CISD-relative sweep window [t-4, t] (Default: 5).
• SWEEP_SWING_LOOKBACK: Lookback used to source prior swing points for sweeps (Default: 20).
• Swing SMT support depends on the local SMT package at /mnt/e/backup/code/Finance/Misc/SMT.

• Python 3.10+
• pandas, numpy, matplotlib, pyarrow