Methodology
How Rheinmarx turns market data into signals, measures what happened afterwards, and decides whether a strategy deserves to publish at all.
Principles
The engine is deterministic: the same stored inputs and the same strategy version always produce the same decision. Decisions use closed bars only. A setup that passes every gate is published automatically, with no human approval, and a setup that fails is never published by hand. Nobody can create, edit or delete a signal or its outcome after the fact.
AI is secondary. After publication, a language model may write a short review that confirms the setup, adds caution or flags a conflict. It cannot create, veto, delay or change a signal, and publication never waits for it.
Timeframes
| Regime | Setup | Trigger | |
|---|---|---|---|
| IDX equities | 1D (IHSG) | 1H | 1H |
| Digital assets | 4H (BTCUSDT) | 1H | 15m |
Factor families
Each asset receives five scores from 0 to 100. Setup quality is their weighted sum.
| Family | What it measures | IDX weight | Digital weight |
|---|---|---|---|
| Trend | EMA alignment and slope, agreement with the higher timeframe | 25 | 25 |
| Momentum | MACD histogram, rate of change against volatility, RSI zone, acceleration | 20 | 20 |
| Participation | Volume against its 20-bar median, recent participation, traded value | 20 | 20 |
| Relative strength | Excess return against the benchmark over short and long windows | 15 | 15 |
| Structure | Volatility inside an allowed band, extension from the fast EMA, stop distance, reward headroom | 20 | 20 |
Gates, in order
- Data quality. Required bars must be complete, fresh and consistent. Unhealthy data fails closed: no signal for the affected asset or market.
- Eligibility. The asset must be in the active universe and meet liquidity and listing-age rules.
- Market regime. Benchmark trend and market breadth must not be risk-off.
- Setup quality must reach the strategy's threshold (80 by default). Scores from 65 are shown as Watch.
- Entry structure. A buy-stop entry above recent highs and a stop below recent lows, bounded between 1 and 3 ATR.
- Reward to risk of at least 2.0, with the target capped at nearby resistance.
- No duplicate. No open signal already exists for the asset.
Every gate is evaluated even after one fails, so the terminals can always explain why an asset is not a signal.
Outcome policy
Signals are tracked on the trigger timeframe. Entry fills when price trades through the entry level within the trigger window; a gap beyond the allowed chase distance is skipped; touching the stop before entry invalidates; no trigger within the window expires. When one bar touches both stop and target, the stop is assumed first. No target is credited on the fill bar. Positions still open at the maximum holding period close as timeout. Results are in R after estimated costs. Every published signal counts, including expired and skipped ones.
Validation and promotion
A new strategy version starts as an experiment and publishes nothing; its qualified decisions are recorded as shadow decisions. Promotion to production requires chronological walk-forward folds (4), a final holdout of the most recent 20% of history run once, at least 100 out-of-sample signals, positive cost-adjusted holdout expectancy, holdout drawdown within 15R, positive expectancy in at least 60% of folds, no single asset contributing more than 35% of results, and passing live/backtest parity tests. Validated versions are immutable; changes create a new version.
Backtest, holdout, paper (shadow) and live results are reported separately and never combined.
Reproducibility
Each signal stores its full evidence, the hashes of its input candles and parameters, the engine versions and the data provider mapping. Any signal can be replayed from stored data; the replay result is shown on the signal page.
Limits
Signals describe mechanical setups, not recommendations to buy or sell. Market data comes from third-party providers and can be revised or delayed. Past results, simulated or live, do not guarantee future results.