Methodology

Each signal blends four independent layers, weighted by how predictive each has been. The split below is the default (Neutral-regime) baseline — on-chain leads there — but the engine re-weights all four per detected regime, so the live mix shifts (e.g. technicals lead in trending markets). The live blend is shown on the Signals page.

Weights are calibrated nightly from a k-fold out-of-fold Brier-skill scan against the agent-decision log; any single layer is capped at 0.40 so no one indicator dominates.

Every signal the engine fires is scored against what price actually did. A signal counts as a hit if it closed in profit over its hold, so the bar to beat is roughly 50% — the rate a coin flip would manage. It is not a promise of returns, and it is not a backtest: these are resolved outcomes.

• Hold — signals resolve on a fixed 14-day hold, or earlier if they hit their target or stop.
• Cumulative return — each signal is marked to market across its real time in the trade, and the daily average is compounded. A backlog of resolutions booked on one day therefore cannot inflate it, and overlapping signals on one asset count once.
• Engine eras — the record splits at 2026-07-04, when the balanced-book scorer replaced a previous one whose book ran ~98% SELL. Both eras stay published; the toggle picks which you are reading.
• The stop/target record is a different measurement — it asks whether price reached the target before the stop, and has a much lower chance bar (~28.6%) that moves with the stop width. It is shown separately for that reason, never blended into the hit rate.
• Tamper-evidence — each completed day is SHA-256 hash-chained, committed to git by CI, and anchored into Bitcoin via OpenTimestamps. The Track Record page has a Verify button that recomputes every link in your own browser.

No fees, slippage or execution are modelled. This is a signal-resolution record, not traded performance.

The backtester runs one deliberately transparent rule over real price history, so you can see the whole thing rather than trust a black box. It is a what-if, not a record of trades that happened.

• Entry — go long when RSI(14) is above your chosen threshold AND MACD is above its signal line.
• Exit — flat when RSI falls below the threshold minus 4, OR MACD crosses back below its signal.
• Data — 4-hour bars over the last 365 days, with a 28-bar warmup before the first trade can fire.
• Benchmark — buy-and-hold on the same asset over the same window, also without fees, so the comparison is like for like.
• No fees, no slippage, no execution. Real trading has all three, and they subtract. Treat the result as an upper bound.

Hypothetical performance. A rule tuned until it looks good on one window is the easiest way to fool yourself — which is why the resolved track record above, not this, is the honest measure of the engine.

Beginner, Intermediate, and Advanced change how much annotation you see — never the underlying data. Same numbers, different depth of explanation. Switch levels in the top bar.

• Prices & volumes — Kraken, Gate.io, Bybit, MEXC; CoinGecko fallback.
• Macro — Yahoo Finance (DXY, US10Y), Fear & Greed index.
• Funding — OKX / Bybit perp funding.
• On-chain proxies — derived from 24h ticker volume / price (free-tier surrogate).

• On-chain tiles are proxies, not Glassnode reads — labeled as such throughout.
• Liquidity coverage is biased to top-cap pairs on the listed venues.
• Composite weights refresh nightly — intraday regime flips can lag by up to 24h.
• Past performance never guarantees future returns. Worst-loss-streak is the metric to watch.

• VYLOS Yield — Flare-first stablecoin yield + arbitrage.
• VYLOS Ground — Tokenized RWA portfolios.
• VYLOS ETF Advisor — ETF-based fiduciary advisor.