Framework
Behavioral Intelligence Framework
How to characterize, monitor, and respond to behavioral patterns in decision systems.
Overview
A decision is not fully characterized by its average outcome. Two strategies with identical returns can exhibit fundamentally different behavioral signatures — one steady and predictable, the other volatile and regime-dependent. Average performance masks the behavioral structure that determines how a decision will perform in conditions it has not yet encountered.
The Behavioral Intelligence Framework provides the architecture for observing, classifying, and responding to behavioral patterns in decision systems. It moves beyond aggregate metrics to characterize how decisions behave under stress, across regime transitions, over time, and relative to each other. Behavioral intelligence is what transforms a collection of qualified decisions into a portfolio whose risks are understood, not just measured.
This framework is diagnostic, not prescriptive. It does not decide what to do about behavioral patterns — that belongs to allocation and risk management. It decides what to see. The value of behavioral intelligence is in surfacing distinctions that aggregate metrics hide, so that downstream systems can make better-informed decisions about resource deployment.
Design Principles
These are governing constraints, not guidelines. Every architectural decision in the framework must be traceable to one of these principles.
Behavior Reveals Structure
The behavioral signature of a decision reveals more about its underlying structure than its parameter configuration.
Two strategies with identical parameters can behave differently because of subtle interactions between entry timing, exit mechanics, and environmental sensitivity. Observing behavior directly — rather than inferring it from parameters — captures these emergent properties.
Patterns Precede Outcomes
Behavioral degradation is observable before it produces measurable outcome degradation.
A strategy does not go from profitable to unprofitable in a single trade. The behavioral signature shifts first — win rate erosion, drawdown pattern changes, increasing fee sensitivity. Detecting these patterns early enables response before capital is lost.
Regime Context Is Essential
Behavioral patterns are only meaningful in the context of the environmental regime in which they occur.
A drawdown during a trending regime has different implications than the same drawdown during a choppy regime. Without regime context, behavioral classification conflates environmental effects with strategy-level degradation.
Behavior Is Evidence
Observed behavioral patterns constitute evidence that should inform qualification, allocation, and learning systems.
Behavioral intelligence is not decorative analytics. It produces structured evidence that other systems consume — qualification uses it for re-assessment, allocation uses it for ranking, and learning uses it for generation guidance.
System Map
Five layers from raw observation to system response. Observations are collected continuously. Patterns are extracted on batch cadence. Classification maps patterns to behavioral states. Monitoring tracks state changes. Response propagates behavioral intelligence to downstream consumers.
Layers & Components
Each layer has distinct responsibilities, inputs, outputs, and ownership. No component spans multiple layers.
Observation
Collects raw behavioral data from execution, simulation, and market context. The foundation layer that all pattern extraction depends on.
Trade Outcome Collector
Records individual trade outcomes with full context — entry conditions, exit mechanics, regime, timing
Equity Curve Tracker
Maintains continuous equity curves for behavioral time-series analysis
Regime Context Annotator
Tags every observation with the environmental regime in which it occurred
Pattern Extraction
Transforms raw observations into structured behavioral patterns using statistical and ML-based methods.
Win Rate Analyzer
Tracks win rate evolution over rolling windows, detecting erosion and recovery
Drawdown Pattern Detector
Classifies drawdown behavior — frequency, depth, recovery time, regime sensitivity
Fee Sensitivity Analyzer
Detects strategies whose returns are eroded by transaction costs
Classification
Maps extracted patterns into behavioral categories that downstream systems can act on.
Behavioral State Classifier
Assigns each strategy a current behavioral state based on composite pattern analysis
Regime Sensitivity Profiler
Characterizes how each strategy responds to different environmental regimes
Correlation Profiler
Maps behavioral similarity across strategies to detect hidden concentration
Monitoring
Tracks behavioral state changes over time. Detects transitions between states and generates alerts.
Degradation Monitor
Detects sustained behavioral degradation — not single bad trades, but patterns of decline
Anomaly Detector
Identifies behavioral patterns that deviate significantly from established baselines
Response
Propagates behavioral intelligence to downstream consumers. Advisory — does not autonomously take action.
Alert Router
Directs behavioral alerts to appropriate consumers — qualification, allocation, learning
Intelligence Publisher
Makes behavioral intelligence available to downstream decision systems
States & Transitions
A decision moves through these states as it accumulates evidence. Each transition is gated — there are no free promotions.
Healthy
Behavioral patterns are within expected ranges. No degradation signals. Performance consistent with qualification evidence.
Entry: All behavioral dimensions within baseline thresholds for current regime
Exits: Sustained pattern deviation -> Degrading · Sudden anomaly -> Anomalous
Degrading
Behavioral patterns show sustained negative trends. Not yet anomalous, but trajectory is concerning. Advisory alerts emitted.
Entry: Rolling metrics show sustained decline (win rate erosion, increasing drawdown frequency, fee-bleeder detection)
Exits: Patterns stabilize -> Healthy · Degradation accelerates -> Anomalous · Regime change explains pattern -> Healthy (with regime annotation)
Anomalous
Behavioral patterns deviate significantly from all historical baselines. Requires investigation. Strong advisory signals emitted.
Entry: Statistical anomaly detected — level shift, trend break, or volatility spike beyond SPC thresholds
Exits: New baseline established -> Degrading (re-baselined) · Anomaly resolves -> Recovering
Recovering
Previously degraded or anomalous, now showing positive behavioral trend. Confidence rebuilding.
Entry: Positive trend reversal after degrading or anomalous state, sustained for minimum observation window
Exits: Full recovery to baseline -> Healthy · Recovery stalls -> Degrading
| From | To | Trigger | Guard |
|---|---|---|---|
| Healthy | Degrading | 6h evaluation batch detects sustained negative trend | Multiple pattern dimensions showing decline over 5+ data points |
| Healthy | Anomalous | Anomaly detector fires | SPC threshold breach on level shift, trend break, or volatility |
| Degrading | Anomalous | Degradation accelerates beyond anomaly thresholds | Pattern deviation exceeds 3-sigma from historical baseline |
| Degrading | Healthy | Patterns return to baseline ranges | Sustained recovery over minimum observation window |
| Anomalous | Recovering | Anomaly resolves, positive trend emerges | Metrics moving toward historical baseline for 3+ evaluation periods |
| Recovering | Healthy | Full return to baseline behavioral ranges | All pattern dimensions within expected ranges for current regime |
| Recovering | Degrading | Recovery stalls or reverses | Positive trend ceases, metrics plateau or resume decline |
Boundary Rules
Boundaries define what crosses between layers and what does not. Every boundary has an explicit failure mode.
Observation -> Pattern Extraction
Individual trade records with full context, equity curve data points, regime-annotated timestamps
Raw market data, tick-level execution details, order book state
Missing observations -> pattern extraction operates on incomplete data, confidence flags lowered.
Pattern Extraction -> Classification
Structured pattern metrics — rolling statistics, trend classifications, fee ratios
Individual trade details, raw equity curves, computation methodology
Insufficient pattern data -> classification returns 'insufficient_data' rather than a false healthy signal.
Classification -> Monitoring
Behavioral state labels with confidence, regime context, comparison to baselines
Classification model internals, threshold tuning parameters, alternative classifications considered
Classification failure -> monitoring retains previous state (conservative). No silent state changes.
Monitoring -> Response (External Systems)
Advisory alerts, behavioral intelligence summaries, structured evaluation entries
Binding directives. Behavioral intelligence advises — it never autonomously pauses, kills, or reallocates.
Alert delivery failure -> logged but non-fatal. Behavioral intelligence is advisory; its absence degrades quality but does not break execution.
Operational Semantics
How this framework operates in practice — cadence, consumers, and staleness rules.
| Dimension | Value | Rationale |
|---|---|---|
| Observation collection | Continuous (every closed position, every snapshot) | Observations are the raw material. Missing observations cannot be reconstructed. Collection is continuous and append-only. |
| Pattern extraction cadence | Every 6 hours (evaluation batch) | Pattern extraction is computationally meaningful only with sufficient new data. 6h batch aligns with qualification assessment cadence. |
| Classification update | Every 6 hours (evaluation batch) | Behavioral state is updated alongside pattern extraction. Consumers read the latest classification on their own cadence. |
| Anomaly detection | Every 6 hours (M6 model, shadow mode) | SPC-based anomaly detection runs in eval batch. Currently shadow mode — scores stored in evaluation inputs, not yet triggering actions. |
| Correlation computation | Every 6 hours (M5 model) | Portfolio correlation matrix recomputed each eval batch. Feeds diversity scoring and allocation intelligence. |
| Alert retention | Indefinite (alert_log table) | Behavioral alerts are historical evidence. They inform learning systems and audit trails. Never deleted. |
| Rolling window size | 20 trades (fee-bleeder), regime-duration-aware (other patterns) | Window sizes balance recency against statistical reliability. Fee-bleeder uses fixed 20-trade window; other patterns adapt to regime duration. |
Knowledge Lineage
Depends On
Introduced In
Expanded In
Modeled By
Investigated In
Influences
Related Research
Methodologies
See this framework implemented in Orqis