VPIN: Measure Informed Trading Probability to Anticipate Liquidity Crises

VPIN (Volume-Synchronized Probability of Informed Trading) estimates the likelihood that order flow in a market contains informed traders. High VPIN means market makers face elevated adverse selection risk — they are likely trading against counterparties who know more than they do. Market makers respond by widening spreads and withdrawing liquidity, which can trigger cascading volatility. VPIN is one of the few indicators designed explicitly to measure this feedback mechanism.

Section 1: Core Mechanics

The foundational concept behind VPIN is adverse selection. Every market maker knows some trades come from informed participants who have private information about future price. When informed trading is high, the market maker consistently loses on those trades — the informed side knows which way price will move. The market maker's only defense is to widen spreads.

VPIN is derived from PIN (Probability of Informed Trading), developed by Easley and O'Hara in the 1990s. The key innovation by Easley, de Prado, and O'Hara (2012): instead of measuring trading in calendar time (trades per hour), VPIN measures in volume time — equal-volume buckets regardless of how many minutes they take to fill.

Formula

Divide the trading day into equal-volume buckets (default: , each bucket = 1/50 of average daily volume).

For each bucket , classify total volume into buy volume and sell volume , where (total bucket volume, equal for all buckets).

Where:

• is the constant bucket size (equal-volume)
• is the absolute order flow imbalance per bucket
• The sum is over the most recent completed buckets (rolling window)

VPIN approaches 1.0 when every bucket is dominated entirely by one side (all buy or all sell). VPIN approaches 0.0 when each bucket has perfectly balanced buy and sell volume — no informed trading.

Bulk Volume Classification — The BVC Method

VPIN's original paper uses the Bulk Volume Classification (BVC) method to split volume without requiring individual trade-level quote data:

Where is the standard normal CDF, is price change during bucket , and is the standard deviation of price changes across buckets. This allows VPIN computation from OHLCV data alone — no tick-level data required.

Parameters

Section 2: Interpretation & Signals

Signal Zones

Entry and Exit Rules

Avoid entry (risk reduction): VPIN > 0.70 and rising = market makers are pulling liquidity. If you must trade, use limit orders only and expect wider fills.

Entry confirmation: VPIN declining from elevated levels (dropping from 0.65 to 0.45) = informed trading subsiding = liquidity returning = normal trading conditions resuming. This is often a good re-entry signal for trend trades that have consolidated.

Directional signal: When VPIN is high and rising, check the cumulative direction of the imbalance (which buckets are buy-dominated vs. sell-dominated). Sustained buy-dominated buckets = likely price rise. Sustained sell-dominated = likely decline.

VPIN as a Volatility Predictor

VPIN's theoretical value is as a forward-looking volatility signal. When VPIN rises sharply, it signals that informed participants are taking large positions — which implies they expect a significant move. The market maker response (spread widening, liquidity reduction) creates the conditions for volatile execution.

Section 3: Pass vs. Live — Real-Time Reliability

On an active stock with 10 million average daily volume, a 50-bucket VPIN updates approximately every 200,000 shares. During the first hour of trading, this might be every 5–10 minutes. During the midday lull, a single bucket can take 30–60 minutes to fill. VPIN is not a high-frequency real-time indicator — it runs in volume time, not clock time.

Section 4: Practical Use Cases

Real example: During the week of 2024-08-05 (yen carry trade unwind), VPIN on SPY futures reached 0.84 on August 5th — the highest reading in 6 months. Informed sellers (likely funds unwinding leveraged positions) dominated order flow. The extreme VPIN reading signaled that the selling was not retail panic but coordinated institutional flow. VPIN began declining on August 6th–7th (0.74 to 0.58), signaling the informed selling was exhausting — SPY subsequently recovered 8% over the next 10 trading days.

Section 5: Pseudo Code

INPUT:
  volume_data[]  # OHLCV bars: (timestamp, open, high, low, close, volume)
  bucket_count = 50  # number of buckets N
  lookback = 50  # rolling window of buckets for VPIN calculation

PROCESS:
  Step 1: Compute average daily volume (ADV) from last 20 trading days
            adv = mean(daily_volumes[-20:])
            bucket_size = adv / bucket_count  # V

  Step 2: Accumulate volume into equal-size buckets
            current_bucket_volume = 0
            buckets = []
            for each bar in volume_data:
                current_bucket_volume += bar.volume
                if current_bucket_volume >= bucket_size:
                    # bucket complete — classify volume
                    dp = bar.close - bar.open
                    sigma_dp = std_dev(price_changes_recent)
                    z_score = normal_cdf(dp / sigma_dp)  # BVC method
                    buy_vol = bucket_size * z_score
                    sell_vol = bucket_size * (1 - z_score)
                    buckets.append((buy_vol, sell_vol))
                    current_bucket_volume = 0  # reset for next bucket

  Step 3: Compute VPIN over rolling N buckets
            recent_buckets = buckets[-lookback:]
            vpin = sum(abs(b[0] - b[1]) for b in recent_buckets) / (lookback * bucket_size)

  Step 4: Generate alerts
            if vpin > 0.70:
                alert = "HIGH_VPIN_RISK"
            elif vpin > 0.50:
                alert = "ELEVATED_VPIN"
            else:
                alert = "NORMAL"

OUTPUT: vpin — float between 0 and 1; alert — risk level string
EDGE CASES:
  - Insufficient history for ADV: use first available 5 days as proxy; flag as estimated
  - Zero-volume bar: skip; do not contribute to bucket
  - Market close gap: start new bucket accumulation at next open
  - Very low ADV stock: bucket_size < 1000 shares = unreliable BVC classification

Section 6: Parameters & Optimization

Standard Configuration

Parameter Impact

Section 7: Synergies & Conflicts

Section 8: Common Mistakes

Section 9: Cheat Sheet