Quantamental Engine · V3
The direction of every signal comes from a 100% deterministic quant engine — no AI involved. It measures momentum over four fixed horizons, applies the strategy validated for that specific asset class, and sizes the position by volatility. The AI layer writes the explanation you read; it never votes on the signal. No black box: here is the exact logic.
Layer 1
The engine reads the daily price history and asks one question on four fixed horizons: is the price higher than it was 21, 63, 126 and 252 days ago? Each answer is a vote of +1 or −1, and their average is the raw momentum score between −1 and +1. Those horizons were chosen before any testing and are never re-optimized — that is what makes every published result out-of-sample by construction.
| Indicator | Condition | Impact |
|---|---|---|
| Momentum 21dShort horizon | Price > price 21 days ago → +1 vote · below → −1 | 1 of 4 |
| Momentum 63dQuarterly horizon | Price > price 63 days ago → +1 vote · below → −1 | 1 of 4 |
| Momentum 126dSemi-annual horizon | Price > price 126 days ago → +1 vote · below → −1 | 1 of 4 |
| Momentum 252dAnnual horizon | Price > price 252 days ago → +1 vote · below → −1 | 1 of 4 |
| Realized volatility 60dPosition sizing | Exposure = target vol ÷ realized vol (capped 1.5×) | sizing |
| MA 200Regime filter (indices & gold) | Above → 100% exposure · below → 60% defensive floor | exposure |
| ADXTrend strength → conviction | > 25 established trend (conv +0.15) · < 20 no trend (conv −0.10) | conv ± |
| ATR (14)Risk levels | SL = 1.5 × ATR · TP = 3.0 × ATR (1:2 reward/risk) | SL / TP |
Signal computation (exact)
Risk levels — ATR(14)
Layer 2
The same momentum score does not lead to the same position everywhere, because asset classes do not behave the same way. Applying one universal model is precisely how systematic engines fail. Each class gets the strategy its own history supports — and nothing else:
Multi-horizon momentum (21/63/126/252d). Shorting is forbidden here: it structurally destroys value on assets with positive drift.
The only asset class where shorting pays: WTI buy & hold is −0.21 Sharpe, the strategy turns it into +0.47.
No timing model beats buy & hold on indices — so we do not pretend to. Exposure is scaled by volatility and the 200-day average to cut drawdowns instead.
Trend applied to total return (coupons included), long or cash. Modest standalone, valuable as portfolio diversification.
A fifth case matters just as much: no signal at all. On the 10 major FX pairs, silver, copper and XRP, no strategy held up — so none is issued. They are shown as NEUTRAL with macro context only.
Layer 3
Two mechanical rules sit on top of the signal, and neither of them can change its direction — only its size. First, volatility targeting: when realized volatility doubles, exposure is halved, which is what keeps drawdowns contained. Second, an event rule: position size is halved on the day of and the day before a high-impact macro event (FOMC, CPI, NFP, ECB, PCE).
The event rule is a calendar rule, not a forecast: we make no prediction about the outcome of the event, we simply carry less risk into it.
Decision thresholds
Layer 4
Once the signal exists, an LLM receives the full quant context (strategy used, momentum per horizon, horizon agreement, target exposure, realized volatility, MA200 position, RSI, ADX, ATR) plus the live macro context — rate spread 10Y-2Y, VIX, and CFTC institutional positioning. Its job is to explain that decision in plain language, in French and English, with bull/bear points and an institutional reading.
What changed in V3: the previous version let the LLM produce its own directional score, merged with the quant score through a conviction-weighted average. That was removed. An LLM opinion on market direction cannot be backtested — the model knows history, so any retroactive test of its past opinions is contaminated. We do not let unmeasurable input dilute a measured signal.
New data sources — including AI-generated features such as hawkish/dovish scores — go into quarantine: stored and dated, but never used in a signal until forward validation proves they add value. Funding rates and the Fear & Greed index went through this process and were rejected on measurement.
Systemic layer
Once all assets are scored, the Sentinel averages every score_quant across the universe to read the systemic risk regime. Here score_quant is the quant score rescaled to a −100 to +100 range, so a +15 threshold corresponds to +0.15 on the raw scale.
Empirical Validation
Nothing in the engine is fitted to the data. The four momentum horizons (21/63/126/252 days) are fixed before testing, which removes overfitting by construction, and each asset class gets the strategy its own history supports. Measured over 2014–2026 with transaction costs included, across 1041+ rolling 42-day windows. 13 assets failed validation and carry no directional signal at all.
Top assets — measured strategy Sharpe
| Asset | Sharpe | Return | Max DD | Positive windows |
|---|---|---|---|---|
| Solana | 1.22 | +387.6% | −20.8% | 35% |
| Bitcoin | 1.20 | +4717.7% | −31.0% | 43% |
| BNB | 0.87 | +575.6% | −27.2% | 35% |
| Ethereum | 0.72 | +264.4% | −27.4% | 38% |
| Gold | 0.71 | +148.1% | −16.2% | 54% |
Honest framing: a Sharpe of 0.8–1.0 on every single asset does not exist for anyone — the strength comes from combining decorrelated signals (1.36 portfolio Sharpe, -10.4% max drawdown). The engine is a trend filter, not a magic shield: it reduces exposure in slow, sustained bear markets but cannot dodge flash crashes like March 2020. Drawdowns above are real and unsmoothed.
These signals are algorithmic research outputs, not financial advice. Stop-loss and take-profit levels are ATR-derived references — always size positions according to your own risk management.