A Transformer-based model that predicts full Student-t distributions of cross-sectional relative returns across 362 cryptocurrencies at horizons of 1-30 days, using only standard OHLCV price and volume features.
Paper: paper/main.tex | Release: v1.0 (checkpoints + data)
| Metric | Value |
|---|---|
| Rank IC (10d) | 0.124 (Newey-West t = 8.9; conservative overlap-adjusted lags still give t > 6) |
| Long-Short Sharpe (gross, horizon-corrected) | ~4.1 upper bound (see Limitations; earlier releases reported 11.8-13.0 due to a 10-day-return annualization error) |
| Win Rate (10-day windows) | 74.7% of overlapping 10-day holding windows |
| Quintile / Decile Monotonicity | 1.000 / 0.988 |
| Breakeven Cost (horizon-corrected) | ~34 bps one-way |
| IC Positive Months | 15/15 |
(a) Rank IC increases with horizon (0.08 at 1d to 0.17 at 30d; longer horizons overlap more heavily). (b) Cumulative overlapping 10-day L/S spreads — indicative of signal strength, not an implementable equity curve. (c) Near-perfect decile monotonicity (Spearman 0.99). (d) Phantom has the highest IC among directional baselines; a naive low-volatility sort achieves a higher raw IC (0.165 vs 0.124) — Phantom's incremental value is concentrated in mid/high-volatility names (see Limitations).
IC positive in all 15 months. Rolling 60-day horizon-corrected Sharpe always > 0 (min ~0.9). Quintile-membership turnover stable at ~38%/day.
- Survivorship bias: the universe is all Binance USDT pairs with status TRADING at data-fetch time (Mar 2026). Pairs delisted at any point 2017-2026 — including during the test window — are excluded retroactively from training, testing, and the cross-sectional demeaning that defines the target. Reported IC and spreads are upper bounds on live performance.
- Overlapping-return annualization: the L/S series is built from 10-day forward returns sampled daily. Earlier claims (Sharpe 11.8-13.0, 925% annualized, 74.7% daily win rate, 336 bps breakeven, 2.53%/day spread) annualized these as daily returns, a ~sqrt(10) Sharpe inflation. Corrected: gross Sharpe ~4.1, ~93% annualized gross, breakeven ~34 bps one-way.
- Implementability: legs are equal-weighted with no liquidity screen; many bottom-quintile pairs have no margin/perp market (the short leg is partly unimplementable); no spread, impact, or borrow costs are modeled. The signal's incremental IC is concentrated in high-volatility, low-liquidity names. Realistic net-of-cost performance is plausibly in the 1-2 Sharpe range.
- Low-vol overlap: the signal is 0.63 rank-correlated with inverse trailing volatility, and a naive low-vol sort has higher raw IC (0.165). Phantom's independent contribution is significant only within mid/high-vol terciles (t = 2.6 / 4.1).
- Inference: Newey-West lags (Andrews rule, 5) under-correct for the MA(9) overlap in the daily IC series; conservative lags reduce the IC t-stat from 8.9 to roughly 6-7. There are ~43 independent 10-day periods in the sample.
- Encoder: 8-layer pre-norm Transformer on 24 patches of 120-day, 6-channel OHLCV input (512d, 8 heads)
- Decoder: 2-layer cross-attention decoding all 30 horizons simultaneously
- Head: Student-t output (mu, sigma, nu) per horizon - 3 parameters, no mixture
- Anti-collapse: Condition dropout (p=0.15) + encoder variance penalty
- Parameters: 31.7M
The model predicts relative returns (asset return minus cross-sectional mean), isolating idiosyncratic signal from shared market factors. This is the key insight that unlocks cross-sectional predictability:
- Absolute returns from OHLCV are unpredictable at any horizon (consistent with weak-form EMH)
- Relative returns contain ranking signal concentrated in crypto (IC = 0.124 on a survivor universe; partially overlapping the low-volatility effect — see Limitations)
- Stage 1: Pretrain on 413 assets (crypto + equities + forex + commodities) with 6-channel OHLCV features
- Stage 2: Fine-tune on 362 crypto assets only - the signal concentrates in crypto
Eight model versions systematically test the design space:
| Version | Change | IC (10d) | Sharpe | Finding |
|---|---|---|---|---|
| v1-v2 | Synthetic SDE pretrain | - | - | Oracle CRPS on synthetic, zero transfer to real |
| v3 | Real multi-asset pretrain | 0.00 | - | Good calibration, no directional signal |
| v4 | Multi-horizon absolute returns | 0.00 | - | Memorizes training, doesn't generalize |
| v5 | Relative returns | 0.09 | ~1.5* | Signal unlocked |
| v6 | + Funding rate, taker buy | 0.14 | ~1.7* | Features don't help; crypto-only helps |
| v7 | 4h bars | 0.10 | ~0.8* | Worse: 99.7% sample overlap, signal is daily |
| v8 | 362 crypto assets | 0.12 | ~4.1* | Breadth improves L/S diversification; IC slightly lower than v6 |
* Sharpe values in this table were originally computed by annualizing overlapping 10-day returns with sqrt(365), inflating them by ~sqrt(10); corrected figures are shown for all versions (pre-correction: 4.6 / 5.5 / 2.6 / 13.0). All are gross, equal-weighted, on a survivor universe.
pip install -r requirements.txt
# Download checkpoint and data from release
# https://github.com/ADnocap/phantom/releases/tag/v1.0
# Evaluate
python scripts/eval/eval_v5.py \
--checkpoint checkpoints_v8/best.pt \
--test_data data/processed_v8/test.npz
# Full publication analysis (transaction costs, baselines, robustness)
python scripts/eval/publication_analysis.py \
--checkpoint checkpoints_v8/best.pt \
--test_data data/processed_v8/test.npz
# Train from scratch (requires data fetching first)
python scripts/data/fetch_crypto_v8.py --workers 6
python scripts/data/build_dataset_v8.py
python scripts/train/train_pretrain.py \
--data_mode v5_real_assets \
--real_data_dir data/processed_v8/ \
--init_from checkpoints_v5/best.pt \
--head_type student_t \
--n_input_channels 6 \
--context_len 120 --max_horizon 30 \
--d_model 512 --n_layers 8 --n_heads 8 --d_ff 2048 \
--batch_size 512 --epochs 50 --lr 3e-4phantom/
├── src/
│ ├── model.py # PhantomModel (encoder + cross-attn decoder + Student-t head)
│ ├── losses.py # NLL, CRPS, energy distance, combined losses
│ ├── features.py # 6-channel OHLCV feature computation
│ ├── real_data.py # Dataset classes for processed .npz files
│ ├── data.py # Online synthetic dataset (v1/v2 only)
│ ├── sde.py # SDE simulators (v1/v2 only)
│ └── btc_data.py # BTC OHLCV fetching
├── scripts/
│ ├── train/
│ │ ├── train_pretrain.py # Main training script (v3-v8)
│ │ ├── train_v6.py # v6 crypto fine-tuning with new features
│ │ ├── train_v7.py # v7 4h-bar training
│ │ └── train_finetune.py # BTC fine-tuning (v1/v2)
│ ├── eval/
│ │ ├── eval_v5.py # Cross-sectional evaluation (IC, Sharpe, coverage)
│ │ ├── publication_analysis.py # Full publication analysis
│ │ └── ...
│ ├── data/
│ │ ├── fetch_crypto_v8.py # Fetch all Binance USDT pairs
│ │ ├── build_dataset_v8.py # Build crypto-only dataset
│ │ └── ...
│ └── slurm/ # HPC job scripts (LaRuche A100)
├── paper/
│ └── main.tex # Paper (ACM sigconf format)
├── notebooks/
│ └── phantom_v5_analysis.ipynb # Interactive analysis
├── plots/ # All figures
├── logs/ # Training CSV logs
├── experiments.md # Full experiment log (v1-v8)
└── CLAUDE.md # Project instructions
- Liu, Tsyvinski, Wu (2022). Common Risk Factors in Cryptocurrency. Journal of Finance, 77(2):1133-1177.
- Fieberg et al. (2024). A Trend Factor for the Cross Section of Cryptocurrency Returns. JFQA.
- Hackmann (2026). JointFM: A Joint Foundation Model for Distributional Prediction. arXiv:2603.20266.
- Nie et al. (2023). A Time Series is Worth 64 Words. ICLR 2023.
- Gneiting & Raftery (2007). Strictly Proper Scoring Rules, Prediction, and Estimation. JASA, 102(477):359-378.