Flagship TS-01 · Crypto Microstructure Forecaster
A 60-minute return forecaster on the Crypto Extended microstructure dataset — funding, order-book imbalance and depth at 5-minute bars, always flat across funding.
Alpha performance report
Walk-forward backtest
10 major coins on the Crypto Extended microstructure dataset — 15m bars, funding, order-book imbalance and depth
An intraday time-series forecaster on the Crypto Extended microstructure dataset: order-book imbalance, funding pressure and short-horizon volume dislocations are fused into a standardised 60-minute return forecast per coin on 5-minute bars. Positions are opened only on extreme post-cost signals and are always flat by the funding boundary — no overnight inventory. In-sample 2023-01-01 → 2024-06-30, validation 2024-07-01 → 2024-12-31, out-of-sample 2025-01-02 → 2026-06-30. Headline figures are net of taker fees and modelled slippage. Demonstration run — figures are illustrative.
Signal quality · shared by both books
Forecast IC
0.0640
IC info ratio
0.540
Hit rate
56.4%
Horizon
12 bars
Instruments
10
OOS 2025-01-02 → 2026-06-30 · $1.00M book · directional L/S vs long-only
Statistical robustness
Deflated verdict
Adjusts the headline Sharpe for multiple-testing and non-normality, so a high in-sample number is not mistaken for a real edge.
Sharpe survives multiple-testing deflation — the edge is unlikely to be luck.
Probabilistic Sharpe
99.5%
Deflated Sharpe
97.8%
Sharpe t-stat
3.95
Book Sharpe (ann.)
1.88
Signal quality score
0.89
SQS components
Institutional checklist
Quality gates
Each gate is an independent pass/warn/fail threshold an institutional reviewer would apply before allocating.
≥ 1.00
≥ 95%
≥ 2.0
≥ −25%
≥ 10 bps
≥ 0.50
Signal analysis
Information coefficient & signal quality
How accurately the model forecasts forward returns — IC, calibration and out-of-sample forecast tests.
Forecast IC
0.0640
IC info ratio
0.540
Hit rate
56.4%
OOS R²
0.043
Calibration β
0.9300
IC by forecast horizon
Information coefficient and hit rate at each forward horizon (bars).
Rolling information coefficient
Cross-sectional IC of the forecast, evaluated period by period.
Forecast study
Forecasting
How accurately the model forecasts forward returns — out-of-sample predictions against realised, calibration, the statistical forecast tests and the residual checks.
Forecast IC
0.0640
IC info ratio
0.540
Hit rate
56.4%
OOS R²
0.043
Calibration β
0.9300
FAS score
0.87
Out-of-sample · predicted vs realised
Standardised forecast vs realised forward return at the model's horizon — both normalised (mean 0, unit variance), so this reads as directional agreement, not price scale.
Forecast residual · BTCUSDT
Forecast calibration
Predicted vs realised standardised return; slope ≈ skill.
Forecast diagnostics
Clark–West
4.70
Directional (PT)
11.40
MZ slope β
0.940
IC t-stat (NW)
6.30
Prob. Sharpe
99.5%
Instruments
10
Residual distribution
Pooled residuals with the fitted normal overlaid.
Residual Q–Q plot
Sample vs normal quantiles — points off the line mean fat tails.
Residual autocorrelation
Autocorrelation by lag with the ±95% significance band.
Residuals show significant autocorrelation — expected here because consecutive forecasts share an overlapping horizon, which is why inference uses Newey–West standard errors.
Per-instrument forecast leaderboard
| Instrument | IC | Hit | OOS R² | Calib β | Obs |
|---|---|---|---|---|---|
| BTCUSDT | 0.0781 | 58.1% | 0.05 | 0.9739 | 148,000 |
| ETHUSDT | 0.0741 | 58.3% | 0.06 | 0.9588 | 144,600 |
| SOLUSDT | 0.0694 | 57.2% | 0.05 | 0.9607 | 141,200 |
| BNBUSDT | 0.0652 | 57.0% | 0.05 | 0.9125 | 137,800 |
| XRPUSDT | 0.0691 | 55.6% | 0.04 | 0.8986 | 134,400 |
| ADAUSDT | 0.0580 | 55.2% | 0.04 | 0.9121 | 131,000 |
| DOGEUSDT | 0.0624 | 54.8% | 0.03 | 0.8766 | 127,600 |
| AVAXUSDT | 0.0559 | 54.1% | 0.03 | 0.8980 | 124,200 |
| LINKUSDT | 0.0461 | 53.6% | 0.03 | 0.8692 | 120,800 |
| TRXUSDT | 0.0552 | 52.9% | 0.03 | 0.8607 | 117,400 |
Portfolio construction
Tradeable books
The signal turned into portfolios: a market-neutral long/short and a long-only variant, each marked to an out-of-sample equity curve.
Out-of-sample equity
Growth of the initial book over the walk-forward test period.
Long / Short
L/STotal return
+23.2%
Ann. return
+15.0%
Sharpe
1.88
Sortino
2.80
Max DD
-3.6%
Win rate
54.5%
Profit factor
1.28
Turnover
62.0×
Drawdown
Underwater curve from the running peak.
Daily return distribution
Histogram with mean, ±1σ band and VaR₉₅.
Rolling Sharpe
Annualised Sharpe over a trailing window.
Rolling volatility
Annualised volatility over a trailing window.
Daily returns
Calendar of realised daily P&L over the out-of-sample window.
Long Only
L/OTotal return
+44.1%
Ann. return
+27.7%
Sharpe
1.14
Sortino
1.70
Max DD
-24.0%
Win rate
51.6%
Profit factor
1.16
Turnover
27.0×
Drawdown
Underwater curve from the running peak.
Daily return distribution
Histogram with mean, ±1σ band and VaR₉₅.
Rolling Sharpe
Annualised Sharpe over a trailing window.
Rolling volatility
Annualised volatility over a trailing window.
Daily returns
Calendar of realised daily P&L over the out-of-sample window.
Head to head
Long/short vs long-only
The same signal expressed two ways — compared on growth, drawdown, the risk-adjusted profile and year by year.
Cumulative return
Growth of each book over the test window.
Drawdown
Underwater curves overlaid.
Risk-adjusted profile
Sharpe, Sortino, Calmar, win rate, profit factor and low-vol, normalised.
Annual returns
Year-by-year return for each book.
Tradeability
Transaction-cost sensitivity
How performance survives realistic trading costs — and the breakeven cost at which the edge disappears.
L/S breakeven
26.0 bps
L/O breakeven
33.0 bps
L/S turnover
62.0×
L/O turnover
27.0×
Net Sharpe vs cost
Annualised Sharpe net of trading cost at each bps tier; dashed line marks breakeven.
Net return vs cost
Total return net of trading cost at each bps tier.
Risk & costs
Risk profile
Tail risk, distribution shape and trading cost intensity for the headline book.
Ann. volatility
7.6%
Daily volatility
0.40%
Skewness
-0.13
Kurtosis
3.5
VaR 95%
-0.64%
CVaR 95%
-0.83%
VaR 99%
-0.92%
Max consec. losses
6
Avg drawdown
-0.86%
DD recovery
21d
Total trades
10,153
Annual turnover
62.0×
Intraday execution
Trade-level statistics
How the intraday book actually trades within each session — turnover, holding time, per-trade economics and the forced end-of-day flatten.
Trades / day
12.4
Total trades
6,758
Avg holding
47 min
Forced EOD-flat
1,036
Avg P&L / trade
+0.019%
Trade hit rate
56.3%
Avg win
+0.108%
Avg loss
-0.083%
Sessions
545
Flat by session close, fresh each day. Positions are opened and closed within the trading session and force-flattened at the close — there is no overnight exposure and each session starts from cash.
Reproducibility
Backtest configuration
Exactly how the run was specified — universe, frequency, walk-forward scheme and capital.
Universe & method
Walk-forward & capital
Execution rules
Minutes to a few hours — always flat across each 8h funding boundary.
Final value $1.23M from $1.00M over 1.49 years · generated 2026-07-14.
How it works
Methodology & next steps
The mechanism behind the signal and the improvements the research pipeline proposes next.
Algorithm
- Build per-coin microstructure features each bar: depth-weighted order-book imbalance, funding-rate pressure, signed volume surprise and short-horizon realised-volatility ratios.
- Standardise features on a rolling window (no look-ahead) and fit a gradient-boosted forecaster of the forward 60-minute return, re-fit monthly walk-forward.
- Emit a standardised forecast z each bar; open a position only when |z| ≥ 1.8 and expected edge clears fees + slippage.
- Exit when |z| < 0.5, on a 3.5σ give-up stop, or at the 8h funding boundary — the book is always flat across funding.
Key features
- Order-book imbalance, funding pressure and volume-dislocation features from the 1m microstructure feed
- Standardised 60-minute forecast per coin, re-fit monthly on a strict walk-forward schedule
- Trades only |z| ≥ 1.8 post-cost signals; force-flat at every 8h funding boundary
- Directional L/S book with a long-only variant; both marked to a daily out-of-sample curve
Known limitations
- Edge concentrates in the top-5 coins by depth; the tail coins contribute breadth, not P&L
- Signal capacity is bounded by top-of-book depth at the 5m horizon
Proposed improvements
Cross-venue order-flow features
Perp flow on secondary venues leads price on the primary venue at short horizons.
Impact. Higher IC at the 15–60 minute horizons.
Adaptive entry threshold
Fixed |z| ≥ 1.8 leaves edge on the table in high-vol regimes.
Impact. More trades per day at equal hit rate.
Jump to section