From Market Tape to Decisions

Each stage transforms raw exchange data into calibrated predictions. Select a node to inspect.

Input

Engine

Core Model

Output

ML-Powered

ML Candidate

Observe

Exchange APIs

OHLCV, OI, Funding Data source

Historical DB

Labels, Snapshots Storage

ingest

Interpret

Outlook Engine

14 factors -> score ML candidate

Gap Registry

FVG lifecycle ML candidate

Feature Builder

ML vectors Feeds ML

transform

Predict

Imminent Model

P(move) calibrated ML

Target Model

Selection Logic ML candidate

Interaction Model

P(fill), P(react) ML

decide

Act + Learn

Discord Alerts

Trigger notify Output

Dashboard

Live state Output

Learning Loop

Feedback -> Retrain ML

Observe

OANDA API

15M XAUUSD Candles Data source

Historical DB

numerian_gold.db Storage

ingest

Interpret

S/R Zone Detector

15M S/R zones ML candidate

Trendline Tracker

Channel detection ML candidate

Regime Classifier

Range / Trend / Breakout ML candidate

Feature Builder

S/R + Trendline vectors Feeds ML

transform

Predict

Imminent Model

P(breakout) + CVS Planned

Target Model

S/R zone or trendline Planned

Interaction Model

Bounce vs Break Planned

decide

Act + Learn

Discord Alerts

Gold notifications Output

Gold Dashboard

Live XAUUSD state Output

Learning Loop

Feedback -> Retrain Planned

Select any node in the diagram to inspect its role, inputs, outputs, and example payloads.

Input

Exchange APIs

Real-time market data ingestion

What it does

Pulls live candlestick data (OHLCV), open interest snapshots, and funding rate info from crypto exchanges. This is the raw signal everything else is built on.

ML Status

Data Source - not ML

Raw data feed. No learned parameters. Feeds downstream ML through feature engineering.

Outputs

candles[] oi_data funding_rate

Example

"close": 104250.5, "volume": 1823.4, "funding": 0.0001

Input

Historical Database

Labeled outcomes and snapshots

What it does

SQLite store containing past predictions, labeled outcomes (did a move happen?), gap snapshots, and training data. The system's memory.

ML Status

Storage - not ML

Persists training labels and snapshots. The system's memory - consumed by ML but not ML itself.

Key Tables

outlook_target_snapshots outlook_move_labels fvg_registry fvg_eval_labels

Engine

Outlook Engine

Market readiness scoring

What it does

Computes 14 factors—volume compression, range shrinkage, coiling, squeeze, OI divergence, aggressor flow, funding rate, session/funding time, plus 4 ML-only signals (OI momentum, crowding, CVD divergence, volume surge). 10 scored heuristically, 4 reserved for ML. Outputs a 0-100 readiness score plus quality gates.

ML Status

ML Candidate - currently rules-based

Factor weights are hand-tuned. Could be replaced with learned weights via gradient optimization or meta-learning on the factor scoring functions themselves.

Inputs / Outputs

candles[] outlook_score factors{} gates{}

Owned by

Imminent Initiative

Engine

Gap Registry

FVG detection and lifecycle

What it does

Detects Fair Value Gaps in price action, tracks their lifecycle (created, tested, filled), and computes F/R/C quality scores. Provides candidate gaps to the Target Model for selection.

ML Status

ML Candidate - currently heuristic

FVG detection uses fixed rules (ATR thresholds, displacement checks). Could train a classifier on historical gap outcomes to learn which gaps are worth tracking.

Inputs / Outputs

candles[] open_gaps[] primary_target frc_scores

Owned by

FVG Initiative

Engine

Feature Builder

ML feature extraction

What it does

Transforms raw data and engine outputs into structured numeric vectors for ML models. Handles missing values, normalization, and maintains schema stability across versions.

ML Status

Feeds ML - deterministic feature engineering

Deterministic code that transforms raw data into structured numeric vectors (32 + 35 dims) feeding the Imminent and Interaction models. Not a learned model itself - schema-stable, null-safe, normalized.

Inputs / Outputs

outlook_factors gap_context feature_vector[] feature_names[]

Tags

schema-stable normalized null-safe

Core Model

Target Model

Primary target selection

What it does

Selects the "Primary Target" from open gaps using distance, quality scores, vacuum bias, and market context. Answers: WHERE will price go next?

ML Status

ML Candidate - currently rules-based

Uses a fixed 45/45/10 weighted rank. Could learn optimal weights from historical target selection accuracy, or use a learned model to directly predict which gap will be visited next.

Inputs / Outputs

open_gaps[] current_price primary_target direction

Owned by

FVG Initiative

Tags

heuristic weighted-score

Core Model

Imminent Model

Move probability prediction

What it does

Calibrated logistic regression that outputs P(large move within H hours). Uses Platt scaling for reliable probabilities. This is the main go/no-go trigger signal.

ML Status

ML-Powered - CalibratedLogReg

3 calibrated logistic regression models (H1/H6/H24) with Platt scaling. Retrained daily at 04:00. This is the primary decision engine.

Inputs / Outputs

feature_vector prob: 0.0-1.0 triggered: bool

Owned by

Imminent Initiative

Tags

calibrated thresholded time-split

Core Model

Interaction Models

Gap behavior prediction

What it does

Two calibrated models predicting what happens when price reaches a gap: P(fill) for complete fills, P(reaction) for confirmed bounces. Helps set expectations on target behavior.

ML Status

ML-Powered - CalibratedLogReg

Two calibrated logistic regression models: P(fill) and P(reaction). Retrained daily at 03:00 from labeled gap-touch outcomes.

Chart Visualization

Realized outcomes are shown directly on the chart as diamond markers on the candle where the fill event occurred:

REJ - Rejected (zone held) BRK - Broke through (zone failed) HOV - Hovering (undecided)

Markers persist for 24h (echo window). Tap on mobile to see full outcome text. A connecting line anchors each marker to the gap zone midpoint.

Inputs / Outputs

gap_context fill_prob reaction_prob chart markers (REJ/BRK/HOV)

Owned by

FVG Initiative

Output

Discord Alerts

Trigger notifications

What it does

When models trigger above threshold, rich embeds are sent to Discord with full context: probabilities, target gaps, outlook scores, F/R/C, and confidence levels.

ML Status

Output - not ML

Consumes ML predictions. Does not contain any learned parameters.

Inputs

trigger_event context_payload

Tags

webhook cooldown: 60m

Output

Dashboard

Real-time state visualization

What it does

Live web interface showing current outlook state, factor scores, gap targets, model probabilities, and system health. Auto-refreshes every 5 seconds.

Chart Features

Gap zones with strength ratings, reach probabilities, F/R/C score chips, dormant zones, echo footprints for recently-filled gaps, and interaction outcome markers (REJ/BRK/HOV diamonds) on fill candles. Mobile drawer shows recently-filled gaps with outcome tags.

ML Status

Output - not ML

Visualization layer. Consumes and displays ML outputs but has no learned parameters.

Inputs

/api/state /api/gaps

URL

num.devircle.com

Feedback System

Learning Loop

Self-correcting backbone

What it does

This is how the system gets smarter. It keeps track of what it predicted versus what actually happened. When it's wrong, it learns from it. When it's right, it remembers why. Over time, the predictions become more reliable.

ML Status

ML-Powered - feedback + retraining

Closes the loop: labels outcomes, evaluates model drift, triggers daily retraining. The self-correcting backbone that makes all ML models improve over time.

Active Loops

Imminent - daily 04:00 Interaction - daily 03:00 Outlook labeling - hourly FVG labeling - hourly

Lifecycle

Observe Label Evaluate Retrain

Input

OANDA API

Primary XAUUSD data provider

What it does

Pulls 15-minute XAUUSD candles via the OANDA v20 REST API. Also provides real-time bid/ask spread data and order book snapshots. This is the raw signal for all Gold pattern detection.

ML Status

Data Source - not ML

Raw market data feed. No learned parameters. Feeds downstream engines through 15M candle aggregation.

Outputs

candles_15m[] spread_data order_book

Example

"close": 2648.35, "volume": 4210, "spread": 0.22

Input

Historical Database

Gold labels, zones, and snapshots

What it does

SQLite store for the Gold pipeline. Contains detected S/R zones, trendline history, regime classifications, labeled outcomes, and training snapshots. Separate from the BTC database.

ML Status

Storage - not ML

Persists training labels, zone history, and snapshots. The Gold system's memory.

Key Tables

sr_zones trendlines regime_snapshots gold_move_labels

Engine

S/R Zone Detector

Horizontal support/resistance identification

What it does

Identifies horizontal support and resistance zones on the 15M timeframe by clustering pivot highs/lows. Tracks zone strength (touch count, recency, clean rejections). This is the primary pattern recognition engine for Gold - equivalent to the Gap Registry in BTC.

ML Status

ML Candidate - currently rules-based

Zone detection uses fixed rules (pivot clustering, ATR-based thickness). Could train a classifier to learn which zones are structurally significant vs noise.

Inputs / Outputs

candles_15m[] active_zones[] zone_strength zone_geometry{}

Owned by

S/R Initiative

Engine

Trendline Tracker

Directional channel detection

What it does

Detects ascending/descending trendline channels on the 15M timeframe. Tracks channel slope, width, number of confirmed touches, and wall proximity. Feeds all three Core Models with directional momentum context.

ML Status

ML Candidate - currently rules-based

Trendline detection uses linear regression on pivot points with touch-count validation. Could learn optimal slope tolerances and channel width from labeled channel-break outcomes.

Inputs / Outputs

candles_15m[] active_channels[] wall_proximity trendline_geometry{}

Owned by

Trendline Initiative

Engine

Regime Classifier

Market state detection

What it does

Classifies the current 15M market state into one of three regimes: RANGE (price oscillating between S/R boundaries), TREND (directional move within a channel), or BREAKOUT (transitioning between regimes). This is the Gold equivalent of the Outlook Engine's readiness detection.

ML Status

ML Candidate - currently rules-based

Uses ADX, Bollinger Band width, and S/R zone proximity to classify regime. Could learn regime boundaries from labeled market-state transitions.

Inputs / Outputs

candles_15m[] active_zones[] active_channels[] regime: enum regime_confidence

Owned by

Imminent Initiative

Engine

Feature Builder

Gold-specific ML features

What it does

Transforms raw 15M data and engine outputs into structured numeric vectors for Gold ML models. 10 feature groups: A (session/time), B (volatility), C (momentum), D (spread/liquidity), E (cross-asset), F (macro/calendar), G (positioning), H (S/R zone geometry), I (trendline geometry), J (regime state).

ML Status

Feeds ML - deterministic feature engineering

Deterministic code building structured numeric vectors from Gold-specific raw data. Groups H (S/R geometry) and I (trendline geometry) are unique to Gold - not shared with BTC.

Feature Groups

A: Session/Time B: Volatility C: Momentum D: Spread/Liquidity E: Cross-Asset F: Macro/Calendar G: Positioning H: S/R Geometry I: Trendline Geometry J: Regime State

Core Model

Imminent Model

Breakout probability + CVS

What it does

Predicts WHEN the current market structure will break. For ranging markets: "Is this S/R range about to break?" For trending markets: "Is this trend exhausting?" Includes the CVS (Cascade Vulnerability Score) sub-gate that detects fragility through a two-gate system.

ML Status

Planned - not yet trained

Will use calibrated logistic regression on Gold feature vectors. CVS sub-score uses a two-gate fragility detection: Gate 1 (market stress) AND Gate 2 (structural fragility).

Inputs / Outputs

feature_vector regime_state prob: 0.0-1.0 cvs_score triggered: bool

Owned by

Imminent Initiative

Core Model

Target Model

Level/boundary selection

What it does

Selects WHERE price will go next. In RANGE regime: identifies the nearest untested S/R zone boundary. In TREND regime: identifies the trendline wall or S/R zone that will be encountered. Answers: "Where is price heading?"

ML Status

Planned - not yet trained

Will use composite ranking of S/R zones and trendline boundaries based on proximity, strength, and regime context. May learn optimal weights from historical target-visit accuracy.

Inputs / Outputs

active_zones[] active_channels[] regime_state primary_target target_type direction

Owned by

S/R Initiative

Core Model

Interaction Model

Bounce vs break prediction

What it does

Predicts WHAT happens when price reaches an S/R zone or trendline wall. P(bounce) for reversals off the level, P(break) for clean breakouts through it. The Gold equivalent of BTC's Fill/Reaction model pair, but framed around S/R and trendlines instead of FVGs.

ML Status

Planned - not yet trained

Will use calibrated logistic regression models for P(bounce) and P(break). Trained on labeled zone-touch and trendline-touch outcomes.

Inputs / Outputs

zone_context trendline_context bounce_prob break_prob

Owned by

S/R Initiative

Output

Discord Alerts

Gold trigger notifications

What it does

When Gold models trigger above threshold, rich embeds are sent to Discord with full context: breakout probability, CVS score, target S/R zone or trendline, regime state, and confidence levels.

ML Status

Output - not ML

Inputs

trigger_event gold_context_payload

Output

Gold Dashboard

Live XAUUSD visualization

What it does

Live web interface showing current Gold state: S/R zones on chart, trendline channels, regime badge, model probabilities, session timing, and macro calendar. Separate page from BTC dashboard.

ML Status

Output - not ML

URL

num.devircle.com/gold

Feedback System

Learning Loop

Gold self-correcting backbone

What it does

Tracks Gold predictions vs outcomes: did the S/R zone hold? Did the trendline break? Was the regime classification correct? Labels outcomes and feeds them back into model retraining. Same concept as BTC's Learning Loop, applied to Gold-specific patterns.

ML Status

Planned - not yet active

Will close the loop: label zone-touch outcomes, evaluate regime prediction accuracy, trigger periodic retraining. Architecture designed, implementation awaiting engine completion.

Planned Loops

Imminent - TBD cadence Interaction - per zone touch Regime labeling - per 15M candle

From Market Tape to Decisions

Model Health Monitor

Learning Systems

Data MapEverything the system ingests, stores, transforms, models, and outputs