      ← ALL WORK
      /
      AI PROJECTS
      /
      NUMIN — PATTERN ENGINE
    

Teaching the chart to remember.

How Punch built a pattern-recognition engine that finds historical market analogues — and projects how the next thirty days of bars may unfold.

QUANTITATIVE ANALYSIS PATTERN RECOGNITION FORECASTING CLIENT — NUMIN · QUANT PARTNER SINCE 2014

        PATTERN B-7 DETECTED · SEEN 212× IN HISTORY (SAMPLE)
        PROJECTION ENGINE · IN ACTIVE REFINEMENT
      

SAMPLE FRAME — SPEC PE-01 · HERO LOOP: CANDLE CHART CLOSE-UP, PROJECTION DRAWS IN · 6–10S

Ten years of asking Numin's data questions.

ONE CLIENT · FOUR ENGAGEMENTS

2014 — IT BEGINS Platform engineering for a quantitative team with big questions.

2018 — PRODUCTION ML Connected models trade the markets. 8 TB of data, 78,531 trades, QC metrics on every subtask.

2019 — CLOUD SCALE Test simulations scaled to hundreds of Google Cloud machines, on demand.

NOW — THE PATTERN ENGINE The chart learns to remember. This page.

Numin — archival trading workstation, charts traced in signal blue

THE MACHINE · THREE MODULES, BUILT IN PARALLEL

01 · B-LEVEL PATTERNS

The spotter.

Scans current-date price bars and flags B-level formations the moment they complete.

02 · CYCLE PATTERNS

The metronome.

Identifies recurring cyclical behavior hiding across years of market data.

03 · PATTERN MAPPER

The memory.

The analytical core — links every detection to its historical precedents and powers the forward projection.

DETECT

Which patterns are present on today's bars?

→

LOOK BACK

Find every prior occurrence in the history.

→

MAP

Track how price behaved over the next 30 days, each time.

→

PROJECT

Model the path the current setup may take.

→

VALIDATE

Compare projections against actuals; refine.

THE INSTRUMENTS · WHAT A DETECTION LOOKS LIKE

FIG. 02 — B-LEVEL DETECTION · THE SPOTTER FLAGS A FORMATION THE BAR IT COMPLETES (SAMPLE GEOMETRY)

FIG. 03 — CYCLE PATTERN · THE METRONOME FINDS THE BEAT UNDER THE NOISE (SAMPLE GEOMETRY)

THE EVIDENCE · PROJECTED VS ACTUAL

FIG. 01 — PATTERN MAPPER PROJECTION VS ACTUAL · STRONG VISUAL CORRELATION ACROSS TESTED PATTERNS (SAMPLE GEOMETRY)
ANALYTICAL PROJECTIONS, NOT INVESTMENT ADVICE. PAST PATTERNS DO NOT GUARANTEE FUTURE BEHAVIOR.

            MATCH DATEPATTERNCORR+30D MOVE
          

            2019-03-12B-70.91+4.2%
          

            2018-08-04B-70.89+2.8%
          

            2021-11-22B-70.87−1.1%
          

            2014-02-18B-70.86+3.5%
          

            2018-06-30B-70.84+0.9%
          

PATTERN MAPPER TABLE · TOP 5 OF 212 MATCHES · SAMPLE VALUES

Numin — three-monitor workstation at night, charts and tabular data

FIG. 05 — OUTCOME DISTRIBUTION · A TENDENCY, NOT A PROMISE — THE RED TAIL STAYS ON THE CHART (SAMPLE)

FIG. 06 — ICLOSE MIDPOINT GRAPH · SUPPLEMENTARY ANALYSIS VIEW (SAMPLE)

WE GRADE OURSELVES · A HABIT FROM 2018

0.96

DETECTION QC · VS ANALYST LABELS

0.98

MAPPING COVERAGE · MATCHES ACCOUNTED FOR

0.87

PROJECTION CORRELATION · TESTED PATTERNS

Every ML subtask gets an independently measured quality-control metric — a discipline we built with Numin's quants in the first engagement, still running in this one. (SAMPLE VALUES — CONFIRM)

THE FEATURE LAB · WHAT THE MODELS LISTEN TO

FEATURE IMPORTANCE · TOP 7 OF 120+ ENGINEERED SIGNALS

i100_close

1.00

ratio_addiff_t0_vol

0.76

close_mov_rel_open

0.75

linreg_rsqrd

0.71

mass_index

0.68

daily_return

0.66

timediff_in_ticks

0.62

FIG. 07 — RELATIVE IMPORTANCE, NORMALIZED (SAMPLE VALUES)

INDICATOR CORRELATION MATRIX

ADI

CMF

EM

FI

OBV

ADI

1.0

.62

.41

.18

.78

CMF

.62

1.0

.71

.44

.23

EM

.41

.71

1.0

.15

.52

FI

.18

.44

.15

1.0

.39

OBV

.78

.52

.52

.39

1.0

            LOW
            
            HIGH

FIG. 08 — WHICH INDICATORS MOVE TOGETHER · HOVER ANY CELL (SAMPLE VALUES)

THE ENGINE ROOM

Four windows into the engine.

What the team actually watches while the system runs — agents competing, features earning their place, patterns surfacing, correlations forming and breaking.

ILLUSTRATIVE ANIMATIONS · STYLED ON THE LIVE DASHBOARDS

Agent values Every simulated trading agent’s value, live on the board — strategies compete head-to-head, and the leaders rotate as new market data lands.

Feature importance The model showing its reasoning — each signal scored and re-ranked as the regime shifts, so the team always knows which features are earning their keep.

Pattern frequency The chart re-drawing as new windows are scanned — how often each pattern appears in history, updating bar by bar as the engine digs.

Correlation matrix Relationships strengthening and decaying in real time — the matrix shifts as correlations form and break across instruments.

WHAT THE CHART KEPT TO ITSELF

Setups that rhyme with history — invisible at human reading speed.

Analysts know the feeling: this chart looks like something we've seen before. Proving it means scanning years of price data for every prior occurrence — impractical by hand, perfect for a machine.

Punch built an engine that detects patterns on the current bars, finds every historical match, and models how price tended to behave over the following 30 days.

“A decade of price history, asked the same question on every bar.”

212HISTORICAL MATCHES, ONE PATTERN (SAMPLE — CONFIRM)

30 dPROJECTED FORWARD, EVERY DETECTION

3MODULES, DEVELOPED IN PARALLEL

WHAT WE BUILT

MODELS✓Real-time pattern-detection engine✓Historical match model✓30-day forward forecast model

INFRASTRUCTURE✓On-prem training cluster (400+ cores)✓Feature engineering✓Three modules, built in parallel

PRODUCT SURFACE✓Correlation matrix✓Projected-vs-actual scoring✓iClose midpoint view

SERVICES PROVEN Forecasting & predictive modeling Feature engineering Custom model development STATUS In production use; modeling under active refinement — and we say so.CASE STUDY numin case study cover

Case study PDF ↓

Punch / Lab NotesISSUE 14

LATEST BUILD

THIS ISSUEInside the models we shipped this month.What we trained, what surprised us, and the eval that let it ship.

PUNCH · AI RESEARCH NEWSLETTERThe AI we’re building next — in your inbox.Research notes from the lab, the models we’re training, and the new projects we’re shipping. A few times a month — no spam, no fluff.★FREE FOR SUBSCRIBERS · THE AI EVAL PLAYBOOK — SEE INSIDE ›

RECENT ISSUESPer-subscriber avatarsYOLO eval at scaleOffline product recognition· JOINED BY 2,400+ ENGINEERS & FOUNDERS

Need work like this? Talk to the engineers who built it

←PREVIOUS — FIRMATEK

Every pole, inspected from the sky.

NEXT — PILOTLY

It watched your cut. It has notes.→