Market regime and signal crowding detection

Not all market conditions are equal. A token that comfortably hits 2× in a bull market may barely move in a bear one. Running the same signal thresholds regardless of market context means over-trading in bad conditions and under-trading in good ones. The MarketRegimeDetector classifies the current state of the Pump.fun market every 10 minutes using observable on-chain signals and publishes the result to Redis — where it is consumed by the ML model, the Genesis Watcher, and the live trader.

The four regime states

Regime	Conditions	Implication
`bull_euphoria`	Graduation rate >8% (2h window), >100 tokens created/hr, >500 SOL/hr volume	High confidence; thresholds can be loosened
`bull_normal`	Graduation rate >3% (2h window), >30 tokens created/hr	Standard thresholds apply
`transition`	Creation or graduation rate diverges >50% from 7-day average	Caution; conditions are shifting, hold thresholds firm
`bear`	Graduation rate below 2% (24h window), below 15 tokens created/hr, below 50 SOL/hr volume	Tighten thresholds; reduce position sizing

transition is the regime that most often precedes a move in either direction. The system applies conservative thresholds during transition rather than guessing direction — a false optimism in a transition that resolves bearish is more costly than a missed opportunity.

Detection signals

The classifier uses 11 observable inputs computed from on-chain data. All inputs are derived from the trades and tokens tables without external price feeds.

Token creation rate

Tokens created in the last 1 hour, last 24 hours, and the 7-day rolling average. Creation rate is the leading indicator — it rises before graduation rate and falls before it.

Graduation rates

2-hour, 24-hour, and 7-day rolling graduation rates. The 2-hour rate captures current momentum; the 24-hour and 7-day rates provide baseline context.

SOL volume

Total SOL volume traded in the last 1 hour and 24 hours. Volume confirms the graduation rate signal — high graduation with low volume can indicate easy graduation conditions rather than genuine demand.

Active wallets

Distinct wallets that have traded in the last hour. High wallet activity with low graduation rate indicates broad participation without quality outcomes — a common bear-market pattern.

Signal hit rates

Average peak multiple and 2× hit rate from live trading over the last 6 hours. This is the most direct feedback signal — it captures whether the current market is rewarding signals with actual returns.

The classification is rule-based, not ML-based. This is deliberate: regime classification needs to be fast, transparent, and robust in edge cases. An ML model would perform similarly on this problem but would be far harder to debug when it misclassifies during novel market conditions.

Confidence scoring

Each regime classification includes a confidence value between 0 and 1.

Regime	Confidence method
`bull_euphoria`	Scales with how far above each threshold the inputs are; maximum confidence at 2× each threshold
`bear`	Scales with how far below each threshold the inputs are; maximum confidence at 50% of each threshold
`bull_normal`	Fixed confidence of 0.6 — defined partly by exclusion from euphoria and bear
`transition`	Fixed confidence of 0.6 — defined by rate divergence from 7-day average

The confidence value is included in the Redis cache and the ML feature vector, allowing downstream consumers to weight the regime signal appropriately when confidence is marginal.

How regime affects the system

ML feature vector
Genesis Watcher
Live trader thresholds

The regime is encoded numerically and included in both the standard and genesis feature vectors:

Regime	Encoded value
`bear`	0
`transition`	1
`bull_normal`	2
`bull_euphoria`	3

This encoding allows the ML model to condition its predictions on market context. A 3× prediction in euphoria requires different supporting evidence than the same prediction in a bear market — and the model learns this distinction from the training data.

The genesis feature vector includes market_regime_encoded as one of its 3 market context features. Tokens launching during euphoria have a measurably different base rate of success than tokens launching during a bear market.The genesis models are trained on historical launches with their regime context embedded, so the model accounts for regime when computing genesis scores. A token with identical launch dynamics gets a different predicted probability depending on the prevailing regime.

The live trader reads the current regime from Redis when evaluating whether to enter a position. In bear conditions, the ML probability threshold is raised to require higher conviction before capital is deployed. In euphoria, the threshold can be relaxed to capture more opportunities.The regime also affects the SignalCrowdingDetector’s penalty severity — crowding penalties are amplified during bear conditions, where marginal signals are more likely to fail.

Redis caching

The regime is cached at market:regime with a 10-minute TTL and includes the full feature snapshot at the time of detection.

Redis key structure

market:regime → {
  regime: "bull_normal",
  confidence: 0.6,
  detectedAt: 1715000000000,
  features: {
    tokensCreatedLastHour: 47,
    gradRate2h: 0.041,
    gradRate24h: 0.038,
    gradRate7d: 0.035,
    solVolumeLastHour: 312.4,
    solVolume24h: 7841.2,
    activeWalletsLastHour: 283,
    signalHitRate2x6h: 0.31,
    avgPeakMultiple6h: 2.1
  }
}

Any service can call MarketRegimeDetector.getRegime() to get the current regime without a database query. The 10-minute TTL matches the reclassification cadence — the cached value is always within one classification interval of current. Every detection cycle also writes a full snapshot to the market_regime_snapshots table for historical analysis and model training. These snapshots enable the ML training pipeline to match historical signals to the regime that was active at the time they fired.

Signal crowding detection

The SignalCrowdingDetector runs every 60 seconds and measures the ratio of tracked-wallet SOL exposure on a token relative to the token’s total bonding curve SOL. When a disproportionate fraction of the curve is held by tracked wallets, the signal is crowded — there are fewer remaining buyers to drive the price.

Crowding ratio

crowding_ratio = tracked_wallet_sol_in_token / token_curve_sol_total

Crowding ratio	Score penalty	Interpretation
< 0.10	0%	Low crowding; ample room for additional buyers
0.10–0.25	10–20%	Moderate crowding; signal quality slightly reduced
0.25–0.50	20–35%	High crowding; significant score penalty applied
> 0.50	Up to 50%	Extreme crowding; signal severely penalised

The crowding ratio is cached in Redis at crowding:<token_mint> with a 60-second TTL and is included as crowding_ratio in the ML feature vector. The live trader also reads it directly when sizing positions — extreme crowding reduces the position size independent of the ML score.

Crowding penalties compound with bear-regime penalties. A moderately crowded signal in a bear market can have its score reduced by up to 60–70% in total, effectively removing it from consideration. This is intentional — crowded signals in bad market conditions have historically poor outcomes.

Detection cadence summary

Component	Cadence	Redis key	PostgreSQL archive
`MarketRegimeDetector`	Every 10 minutes	`market:regime`	`market_regime_snapshots`
`SignalCrowdingDetector`	Every 60 seconds	`crowding:<mint>`	Written to signal record

Genesis Watcher

How market regime is encoded as a genesis feature and conditions genesis model predictions.

ML features

Where market_regime_encoded and crowding_ratio appear in the full 68-feature standard vector.

Live trader strategies

How regime-aware threshold adjustments are applied in strategy selection and position sizing.

Pipeline phase 3

The broader Phase 3 context: ML inference, anti-signal emission, and all advanced intelligence services.

Getting Started

The Pipeline

Intelligence

ML System

Live Trader

Market regime and signal crowding detection

The four regime states

Detection signals

Token creation rate

Graduation rates

SOL volume

Active wallets

Signal hit rates

Confidence scoring

How regime affects the system

Redis caching

Signal crowding detection

Detection cadence summary

Genesis Watcher

ML features

Live trader strategies

Pipeline phase 3

Build docs developers (and LLMs) love

Getting Started

The Pipeline

Intelligence

ML System

Live Trader

Documentation Index

​The four regime states

​Detection signals

Token creation rate

Graduation rates

SOL volume

Active wallets

Signal hit rates

​Confidence scoring

​How regime affects the system

​Redis caching

​Signal crowding detection

​Detection cadence summary

​Related subsystems

Genesis Watcher

ML features

Live trader strategies

Pipeline phase 3

Build docs developers (and LLMs) love

The four regime states

Detection signals

Confidence scoring

How regime affects the system

Redis caching

Signal crowding detection

Detection cadence summary

Related subsystems