Signals dashboard

Quant screens across 150+ cryptocurrencies and 290 US equities. Up to 32 candidate evidence sources are measured, but an unknown or unproven source receives zero live weight. A direction, timeframe, and range are published only when the asset class and the exact asset/side/score/horizon setup beat their measured no-skill baselines on independent outcomes. Otherwise the page says withheld. Analysis syncs hourly; price and 24h change tick live in between.

The confluence model. Every asset is evaluated by up to 32 candidate techniques. They are not assumed statistically independent. Each raw vote is logged for learning, but unknown, weak, or baseline-level techniques receive zero live weight. A board direction survives only when class-level evidence and the exact current asset/side/score/horizon calibration both have enough independent periods and a conservative edge over the measured no-skill baseline.

The 32 techniques. Multi-horizon momentum alignment; Wilder RSI(14) regime and direction; MACD(12/26/9) histogram level and direction; moving-average stack (SMA20/50/200, computed from real daily bars for both equities and crypto); Bollinger %B with squeeze-and-expansion detection; stochastic (14,3) crosses; Donchian 20-bar breakout or breakdown proximity; volume confirmation versus baseline; on-balance volume trend; swing structure of higher-highs and higher-lows; a momentum divergence proxy that flags new price extremes without RSI support; a volatility regime read separating coiled compression from climactic expansion; a reversal-pattern read (below); how long an asset has been coiled at its own long-run high or low and whether it's decoupled from the broader market (below); a seasonal-analog read comparing the current pattern against the asset's own history one or more years back (below); a valuation-or-positioning layer, positioning now weighted by this asset's own funding-rate percentile once enough of its own history exists rather than a fixed global threshold; a Fibonacci retracement read off the asset's most recent swing, direction-aware and never firing without independent confirmation; open interest relative to this asset's own recent history, paired with price direction to separate real participation from a thin, untrusted move; a time-of-day and day-of-week behavioral profile (UTC, New York, London, and Tokyo session hours — which alone captures midnight ET and the NYSE's 9am/4pm hours — and day of week), learned per asset once a slot has real sample depth and a real effect size; market sentiment (Fear & Greed for crypto, VIX's position in its own recent range for equities, pooled with per-asset community/news sentiment where available); a swing-timing read that separately learns what time of day this specific asset's own daily high or low tends to land, firing only when that proven timing pattern and the asset's current price position both confirm; a hack/exploit-severity read that turns a recent, matched security incident from a public hacks tracker into a bearish signal sized to the dollar loss relative to this asset's own market cap, decaying over roughly two weeks; a cross-asset and cross-sector lead/lag read that looks up whichever other assets or curated crypto-sector composites (DeFi, layer-1s, layer-2s, governance tokens, gaming/metaverse, meme, and other baskets) — in either asset class, including the dollar, gold, oil, and the 2-year/10-year Treasury yield spread — have proven, over the full historical archive, to predict this one's moves some number of days later; for crypto, sustained capital flowing into or out of a matched DeFi protocol's on-chain total value locked, paired with price direction before it fires; options-implied volatility (Deribit's DVOL for Bitcoin and Ether, the front-month options chain for equities) at an extreme relative to that asset's own history, a contrarian fear-or-euphoria read that only fires alongside a genuinely stretched price; for equities, an earnings-calendar awareness read that never votes a direction, only flags elevated gap risk and pulls down conviction accordingly, whenever a stock's next reported earnings date falls inside a call's own expected timeframe; a key support/resistance break read that only counts a level once price has reversed off it more than once, sized by how far this asset has historically moved in the 24 hours after that same kind of break; an accumulation/distribution read that looks for a genuinely coiled range — tightening Bollinger bands or realized volatility well under its own baseline, with price itself still flat — and asks which way on-balance volume is quietly leaning inside it, before that lean shows up in price at all; for crypto, a broad-market-outlier read that compares this asset's own 7-day move against the whole tracked crypto market's, voting only when the asset is genuinely decoupled — moving well beyond, or opposite to, what the market itself is doing — not when it's simply riding the same wave everything else is; and, for crypto, a yield-curve read: the one candidate, out of nineteen tested against the full historical record of every major crypto breakout and breakdown, that actually held up independently in both halves of that history — the 2-year/10-year Treasury spread moving more negative over the preceding five days measurably precedes a crypto breakdown, and only that direction, since the mirror case for breakouts did not hold up the same way.

Reversal detection. A separate read from plain RSI level: it looks for RSI having actually bottomed or topped over the last ~10 bars and turned back, confirmed by at least one independent signal (a stochastic cross, a Bollinger band extreme, swing structure, on-balance volume, or the divergence proxy) — it never fires on RSI alone. Market-wide sentiment adds confidence on top when it lines up: extreme fear on the Fear & Greed index for a crypto bottom, or VIX sitting high in its own recent range for an equity bottom (and the mirror image — extreme greed or a complacent VIX — for tops).

Dwell time and market correlation. Real 52-week (or as much history as exists) highs and lows, and specifically how many days an asset has been sitting within a few percent of one, not just whether it currently is — a fresh one-day touch and a multi-week base at the same level are different setups. Long dwell at a low is read as stored energy for a bounce, the mirror at a high for a pullback, and it carries extra weight when the asset has also decoupled from its usual correlation with the broader market (BTC for crypto, SPY for equities) over the last 30 days, since a move happening on its own is a different setup than one just riding the market. This is a starting assumption, not a fixed rule — the adaptive weighting above corrects it per asset from what actually happens next.

Seasonal analogs. Where an asset has enough of its own history, the current pattern over the last ~90 days is compared against the same-length window roughly one, two, or more years back, using the same correlation math as the market-correlation read above but against the asset's own past. A real resemblance has to clear a fairly high bar (a correlation of at least 0.5) before it counts at all, since only a handful of candidate years exist to compare against and a looser bar would just be fitting noise. When one clears that bar, what happened in the days right after that historical analog becomes a genuine, data-grounded forward hint. In practice this only applies to equities: CoinGecko's free tier caps crypto history at 365 days, which isn't enough to compare against even one year back, so this abstains for every crypto asset rather than reaching for a shorter, less meaningful comparison.

The valuation layer. For equities, Wall Street consensus mean price targets and recommendation ratings: trading well below a buy-rated consensus target votes bullish, trading above the consensus target votes bearish. Trefis does not publish a public API, so consensus targets stand in for model-based estimates; site operators can supply Trefis or other model values through a server-side override, in which case the payload labels the source. For crypto, the layer uses perpetual futures funding rates (crowded positive funding on a parabolic move votes bearish, skeptical funding during an uptrend votes bullish) and trending-list crowding.

Adaptive weighting. Every raw vote is logged, then evaluated at an exact target time. Repeated hourly versions of a 24-hour or 7-day forecast are not counted as separate trials: only non-overlapping outcomes enter the append-only ledger. Aggregate updates and the ledger flag commit in one transaction, so a retry cannot learn twice. A technique earns positive weight only when its conservative record beats the relevant baseline; coin-flip, weak, missing, and negatively skilled evidence stays at zero live weight while continuing to learn.

Leading vs. lagging, and the expected timeframe. Technique labels describe intent, not proof. The displayed timeframe comes only from this asset's independent 24-hour versus 7-day outcomes for the active, positively skilled evidence. If neither horizon clears its baseline, the timeframe is withheld. If both qualify but are not distinguishable, the model reports only a broad 1–7 day uncertainty window and does not permit that ambiguity to become a trade call.

Expected range, top, and bottom. A displayed band is an empirical move interval for the validated horizon, anchored to the model's timestamped reference price. It is not an exact top, bottom, target, or stop. Volatility-only fallback bands may still be logged for calibration, but they are not shown as trade metrics. If the exact current setup lacks a reliable historical horizon or range record, both fields are withheld rather than manufactured.

Which indicator this asset leans on. "Leans on X (n%)" under an asset's name names whichever technique has, on its own, the strongest individually-proven track record for that specific asset — some assets really are better predicted by one kind of signal than another, and this surfaces that once a technique has enough of its own scored history to say so, using the same adaptive-weighting data above.

Prediction track record. Directional accuracy, range containment, reversal outcomes, and intraday results remain separate metrics; a high-containment range can no longer inflate a directional hit rate. The list requires a conservative directional edge over the asset-class baseline and enough independent periods. Historical skill does not automatically validate today's setup, which must pass its own calibration gate.

Data. CoinGecko free API for the top 100 coins by market cap with real daily price and volume history per coin, plus global stats and trending (stablecoins and wrappers excluded), alternative.me Fear & Greed, Bybit linear perp funding rates, Yahoo Finance daily OHLCV and Yahoo analyst estimates. Free feeds can lag or drop symbols; the feeds counter in the status bar shows current coverage, and any technique without data simply abstains rather than guessing.

Refresh mechanics. A scheduled job rebuilds the full payload — scores, ranges, horizons, every technique call — hourly and writes it to this page's cache; every visit reads that cache, so the page itself never runs the engine. This page also re-checks every 10 minutes so a new hour's data appears without a reload. Price and 24-hour change are the one exception: this page separately polls a lightweight endpoint roughly every 20 seconds for a live tick, straight from CoinGecko and Yahoo, without touching the hourly analysis — so the number in the Price column can move between rebuilds, but the score, range, and every technique read next to it stay fixed until the next hourly rebuild.

What the scores are not. A score of 70 is not a 70% probability. Only the separately labelled conservative win estimate is calibrated, and even that is uncertain. News, sentiment, earnings, security incidents, positioning, and macro data are incomplete third-party observations, not omniscience. Missing or stale inputs abstain. The automatic research engine tests calendar, event, cross-asset, regime, and strategy rules with family correction, disjoint walk-forward folds, explicit costs, drawdown, and post-discovery confirmation; it does not publish a pattern merely because an in-sample simulation looked profitable.