Crypto Market Microstructure Research
Automated Analysis Using Kaiko Order Book Data (2017–2023) · Cornell University
20
Studies Completed
5
Exchanges
4M+
OB Files Processed
3
GPU Tasks
I. Order Book Structure
R001: Cross-Exchange Order Book Asymmetry
Kaiko OB Binance · Coinbase · Kraken · OkEX · Bitstamp | 2017–2023
Key finding: Binance is consistently ask-heavy (seller pressure), while Coinbase is bid-heavy (buyer pressure). Kraken is the most symmetric.
R017: Order Book Depth Profile
Kaiko OB Binance · Coinbase · Kraken | BTC-USD 2023-04
Key finding: Coinbase bid depth (94 BTC) > ask depth (70 BTC) — retail accumulation signal. Kraken shows asymmetric asks (183 BTC).
R007: Order Book Resilience After Shocks
Kaiko OB Spread recovery analysis | ±20% shock detection
Key finding: Kraken recovers fastest (avg 2 steps), suggesting tighter market-making. Binance recovery is slowest.
R020: Multi-Asset Liquidity Comparison (Binance)
Kaiko OB BTC · SOL · DOGE · AVAX · MATIC | 2023
| Asset | Spread (bps) | OB Depth (USD) |
|---|---|---|
| MATIC | 0.9 | $8.7M |
| DOGE | 1.3 | $7.2M |
| SOL | 4.75 | $11.0M |
| AVAX | 5.65 | $4.4M |
Insight: MATIC has tighter spreads than many larger-cap tokens, reflecting mature L1 market-making.
II. Price Discovery
R010: Lead-Lag & Hasbrouck Information Share
Kaiko Trades Cross-exchange Granger causality
Surprising result: Bitstamp leads BTC price discovery, not Binance. Coinbase is most lagging. Suggests informed traders prefer Bitstamp's lower-surveillance environment.
R004: Kyle's Lambda (Price Impact)
Kaiko OB OFI regression per snapshot
Key finding: Binance has lowest Kyle's lambda (most liquid). Bitstamp has highest (illiquid = larger impact per unit flow).
R013: Order Flow Imbalance → Price
Kaiko OB OFI beta regression, R² analysis
| Exchange | OFI Beta | R² | p-value |
|---|---|---|---|
| Kraken | 0.000170 | 1.5% | <0.001 |
| Binance | 0.000101 | — | <0.01 |
| Coinbase | — | — | not sig. |
III. Cross-Market Analysis
R011: Futures vs Spot Basis
Kaiko OB Binance Futures vs Spot | Monthly analysis
Key finding: Basis peaked at 823 bps in April 2021 (99% annualized) — extreme bullish sentiment. January 2021 was only 5 bps. The basis serves as a real-time sentiment thermometer.
R018: Spot vs Futures Depth Ratio
Kaiko OB Futures/Spot depth comparison
Key finding: Futures depth is 2–3.4× spot depth, confirming derivatives dominate crypto price formation.
R006: Crypto Stocks vs Traditional Market (WRDS CRSP)
WRDS COIN · MSTR · MARA · RIOT vs SPY · QQQ · GLD · TLT
| SPY | QQQ | GLD | TLT | |
|---|---|---|---|---|
| COIN | 0.49 | 0.52 | 0.06 | ~0 |
| MSTR | 0.50 | 0.55 | 0.09 | ~0 |
| MARA | 0.49 | 0.53 | 0.06 | ~0 |
Insight: Crypto equities behave as high-beta tech stocks, not safe havens. Zero correlation with gold and bonds.
IV. Temporal Patterns
R009: Intraday Periodicity (FFT Decomposition)
Kaiko Trades FFT of hourly volume
Key finding: Dominant periods: 24h, 4.8h, 3h. Binance peaks UTC 10 (Asia), Kraken/OkEX peak UTC 16 (Europe/US overlap).
R019: Spread Convergence Trend (2017–2023)
Kaiko OB Monthly median spread evolution
Preliminary: Binance latest median spread 8.85 bps, Coinbase 2.65 bps. Multi-year convergence analysis in progress.
V. Deep Learning (V100 GPU)
GPU01: Convolutional Autoencoder — Market Regime Discovery
PyTorch Conv1D AE → 16-dim latent → K-means
5 market regimes identified: bullish (44%), bearish (43%), mean-reverting (13%).
GPU02: Ledoit-Wolf Covariance → Min-Variance Portfolio
WRDS+GPU 12 assets (crypto equities + indices) | Ledoit-Wolf shrinkage
Shrinkage = 0.007. Key correlations: QQQ↔SPY 0.946, COIN↔MSTR 0.746, MARA↔MSTR 0.767.
GPU03: Transformer Price Predictor
PyTorch 4-exchange × 64-step | 156K params
Test accuracy: 50.68% (barely above random). Confirms short-term return prediction is extremely hard even with multi-exchange attention.
Methods & Data
- Data: Kaiko consolidated order book (OB10-v2), 58 exchanges, 2017–2023. WRDS CRSP daily stock file.
- Infrastructure: Cornell research3 (112-core), research1 (88-core + 2× Tesla V100 32GB), BioHPC (SLURM cluster).
- Libraries: Python 3.12, pandas, numpy, scipy, matplotlib, PyTorch 2.5, scikit-learn.
- Automation: ComeWealth agent — autonomous multi-server research orchestration with parallel task execution.