Cival Pattern Lab — Reading the Indicator

Pattern Lab is a research instrument. On every closed bar it asks one question: did the rule you chose match? If it did, it marks the bar, simulates the trade that would have been taken at the next bar's open, follows that simulated trade bar by bar to its stop, target or time limit, and adds the result to a scoreboard that is split into train, validation and test periods. The point is to find out whether a rule survives costs and unseen data before you risk money on it.

Everything below is taken from the shipped plugin (v1.0.0). Defaults are the ones the indicator starts with.

1. What is drawn on the chart

There are no zones, lines or volume bubbles. Settings for those (Maximum visible zones, Show touched levels, Show bubble volume numbers, Bubble number color, Compact volume numbers) have no effect in this product.

An arrow is drawn for every match, including a match that happens while a simulated trade is already open. Only one simulated position is allowed at a time, so the number of arrows is usually larger than the number of trades in the scoreboard (in the demo: 111 arrows, 78 completed trades and 1 pending, section 11).

Nothing is drawn on the forming bar. A bar is judged when it has closed; the arrow appears when the next bar starts.

2. Gates every pattern must pass

Let b be the bar that just closed, p the bar before it, and prior the Prior lookback bars (default 20) bars before b.

Volume here is bid plus ask volume from the footprint. Delta is ask volume minus bid volume (aggressive buying minus aggressive selling). Close location is (Close − Low) / (High − Low): 0 means the bar closed on its low, 1 on its high. The 0.65 close-location threshold below is fixed in the plugin; it is not a setting.

3. The six patterns

After the gates, exactly one of these conditions decides the side: +1 is a long hypothesis, −1 a short hypothesis.

Notes that matter when you interpret a match:

• The delta gate checks size, not direction, for Engulfing and CandleSequence. A bullish engulfing bar with a negative delta of at least 20% of its volume still passes.
• Absorption is a name for a pattern, not a measurement of absorption. It means "net aggressive flow one way, close location the other way" on one bar. The Lab does not look at what happened at individual prices. Like every family here it is an evidence-based hypothesis about one bar; it does not reveal anyone's hidden positions or orders.
• A CandleSequence with a character other than U, D, N, or a length outside 2–20, never matches and shows no warning. The panel does not print the sequence either; it only prints the pattern name.
• Require directional POC migration (default off) adds one more gate to any pattern: the POC of b (the price row with the most volume; ties go to the lower price) must be higher than p's POC for a long, lower for a short.
• The default settings (DeltaReversal, 20 bars, volume ≥ 1×, |delta| ≥ 20%) and the default DDDUU sequence are examples, not tuned recommendations.

4. From signal to simulated trade

1. Signal bar. The bar that matches is marked. It is closed; nothing about it will change.
2. Entry. The simulated trade enters at the open of the next bar. It is never filled inside the signal bar: a signal on bar 10 cannot enter at bar 10's own open or at any earlier price, so the first possible entry is the open of bar 11.
3. Levels. Stop distance = Stop distance (ticks) × tick size (default 20 ticks). Target distance = stop distance × Target / stop multiple (default 2). For a long, the stop is below the entry and the target above; for a short the reverse.
4. Checked on every bar from the entry bar on (the entry bar counts as bar 1):

5. Both levels inside one bar. Intrabar order is unknown, so the Lab always assumes the stop was hit first and counts the trade as Ambiguous.
6. Net result = side × (exit − entry) − Round-trip costs in price units (default 0). Results are in price units (for a BTC perpetual quoted in dollars: dollars per 1 BTC), not in ticks and not in your account currency.
7. One position at a time. While a simulated trade is open, further matches are marked with an arrow but do not become trades.
8. Censored. If the indicator skips a bar while a trade is open (a hole in the data), that trade is dropped and counted as censored; it is not guessed.
9. Pending. A trade that has not finished at the end of the loaded history is counted as pending and is not added to the statistics.

MAE / MFE. Maximum adverse and favorable excursion are measured from the entry using the full high-low range of each bar, including the bar where the trade exits. Because the real order of highs and lows inside a bar is unknown, they are bounds, not exact excursions. They appear only in the CSV (mae_bound, mfe_bound).

5. Holdout periods: train, validation, test

Set Validation begins (chart time) and Untouched test begins (chart time) to split the history by date. A trade belongs to a period according to when it was known to have finished:

The exit time the Lab uses is the open time of the bar after the exit bar, the earliest moment the result could be known. A trade that exits on the last bar before a boundary is therefore treated as crossing it.

ATAS may display an unset date as 01/01/0001 00:00:00; that means not set. The dates must be in chronological order. If the Test date is set but not later than the Validation date (or Validation is not set), the panel prints INVALID: test must follow validation and no new signals are marked or simulated until the dates are fixed.

Until the dates are set, the panel prints Set holdout dates before judging results, Validation not configured and Test not configured. If you set only Validation, everything after it is Validation. Tip: leave the Test date empty while you explore, so that the Test: line cannot tempt you (see 04-Trading-Playbook.md).

6. The panel, line by line

The panel has nine lines. By default only the first three are visible; untick Compact panel (title + two lines) (group Information panel) to see all of them. When an export error occurs, a line with the error is inserted as line 2.

How the numbers are computed

A group with only a handful of trades produces numbers that are anecdotes, not evidence. The demo's 24 to 27 trades per period are already on the small side.

7. CSV exports

Set CSV export folder (blank disables) to write four files. Use a different folder for each chart or timeframe, because a second instance writing to the same folder overwrites the files.

Column notes:

• partition is train, validation, test or boundary-excluded, using the rules in section 5. With no dates set every trade is train.
• side is 1 (long) or -1 (short). ambiguous is True or False. reason is Stop, Target, Stop gap, Target gap or Time.
• net_price_units is after costs. mae_bound and mfe_bound are full-bar bounds (section 4).
• entry_time is the open time of the entry bar; exit_time is the time the result became known (open of the bar after the exit bar).
• bars.csv: volume and delta are footprint totals (bid + ask, and ask − bid); poc is the bar's highest-volume price; complete is True only when the bar's footprint rows exist and match the candle volume within 1%. Bars without a footprint show volume 0, delta 0, complete False. The file keeps at most Maximum bars back bars.
• signals.csv: time is the signal bar's open time; available_time is the open time of the next bar, when the signal bar's close is known; volume is the marker size and is always 1 for Pattern Lab. Signals older than the last 1,000 retained closed bars are not written to this file.
• All times are in the chart's time base, written in ISO 8601. Numbers use a dot as decimal separator.
• Files are snapshots. Changing a setting recalculates the history and rewrites them, so copy files you want to keep. After you set a folder, files may appear only when the next bar closes or when history is next recalculated; if you do not want to wait, change any calculation setting and change it back.

8. Alerts

Alerts on closed signal / level touch (default off) plays the Alert sound (default alert1) when a bar closes with a match, with the text Cival LAB <Pattern>: <price>. In Pattern Lab an alert means "the rule matched on the bar that just closed". It is an observation for research or replay, not an instruction. Recalculating the history does not fire alerts for old signals.

9. Display and session settings

Setting changes that alter the calculation (pattern, filters, stop, target, holding bars, costs, dates, Maximum bars back) recalculate the whole history. Each change is another trial of the rule, which matters for honest testing (see the playbook).

10. History revisions and what "closed bar" means here

Signals are decided only from bars that have closed; nothing is looked up from the future. However, Pattern Lab recalculates its entire history whenever you change a calculation setting and whenever the data provider revises an already-processed closed bar. Marks and statistics can therefore change after a recalculation. That is a property of the data and of your settings, not a live "repaint". Source corrections, a different history prefix (loading more or fewer bars), changes to any calculation setting and a change of timeframe can all legitimately recalculate the results, so a screenshot taken today may not match a chart opened next week. Panning, zooming or moving the cursor changes only what is drawn, never the underlying closed-bar results.

11. Worked example — real numbers from the demo

Source. The store demo: Hyperliquid BTC perpetual, 5-minute bars (tick size 1.0) built from trades recorded by the Cival order-flow recorder, 17–21 October 2026, 1,380 closed bars (about 4.8 days). It was produced by the shipped plugin. Settings: DeltaReversal with the default gates (volume ≥ 1× the prior-20-bar mean, |delta| ≥ 20%), Stop distance (ticks) 100, Target / stop multiple 2, Maximum holding bars 24, Round-trip costs in price units 24 (about 0.03% of price), Validation begins 19 Oct 00:00, Test begins 20 Oct 12:00. Raw output and settings file: demo\out\pattern-lab.json, demo\settings\pattern-lab.json.

A trade that worked. The 5-minute bar opening 15:45 UTC on 17 October (open 81,434, high 81,434, low 81,372, close 81,374) traded 137.0149 BTC against a prior-20-bar mean of about 22.0 BTC (6.2×). Its delta was −106.0531, or −77.4% of its volume. The previous bar had delta +7.3557 and closed at 81,435. So: previous delta positive, this delta negative, close below the previous close, and the close at the very bottom of the range (close location 0.03, far under the 0.35 limit). The Lab marked ▼ at 81,434 (the bar's high). The simulated short entered at the next open, 81,374; stop 100 units higher at 81,474; target 200 units lower at 81,174. It exited at the target on the 16:55 bar: gross +200, minus the 24 cost = +176.

A trade that did not. The 18:40 bar (high 81,529, low 81,501, close 81,507) traded 103.5589 BTC (about 2.2× its prior mean) with delta −77.3548 (−74.7%) after a previous bar with delta +5.7342. The Lab marked ▼, entered at the next open 81,507, and the 19:40 bar's high reached the stop at 81,607: gross −100, minus 24 = −124.

The scoreboard (panel values; the profit-factor column for the partitions is computed from the trade list in the same raw output using the panel's own formula):

Other numbers from the same run: maximum drawdown 2,931 price units; net result over all 78 trades −2,500 price units; 52 stops, 22 targets and 4 time exits; 0 ambiguous bars; 0 censored; 1 pending; 111 arrows in total.

The arithmetic of why. With a 100-unit stop and a 24-unit cost, a stop-out loses 124 and a target wins 176. Ignoring time exits, that needs about 41.3% winners just to break even (124 ÷ 300). The rule won 32.1% overall. Even the train period, which came closest at 41.7%, still had negative expectancy.

Verdict. This rule lost money after costs in all three periods, including the test period it never saw. It is a rule to reject, and finding that out cost nothing. This is one rule on one market over a short sample, so the honest reading is "no evidence of an edge here", not "this pattern can never work". It also shows the correct response: do not go back and tune the test period away (see 04-Trading-Playbook.md).

12. What Pattern Lab can and cannot know

Unavailable data is not a measured zero: with no footprint rows, too little history or a rule that never matched, the panel simply shows no signals. That says nothing about whether the rule works.

Every change to a setting after you have seen a result is another trial. Many parameter trials create selection bias: with enough trials, some rule will look good by chance. A zero-cost default is not realistic performance, and open simulations are never silently completed. Pattern Lab has no way to count your trials for you; keep your own log.