Signal Engine V5 Is Live: Eight Scored Components, Cross-Exchange CVD Confirmation, and Why We Killed Single-Source Signals

Signal engines in crypto get marketed like magic. The truth is more boring and more useful. A signal is only as good as the components that feed it, the way those components are weighted, and the discipline applied when none of them agree. Buildix Signal Engine V5 is now live across all 530+ pairs we monitor on Hyperliquid, Binance, Bybit, OKX, and dYdX, and it is the version where we stopped trying to make every minute look interesting.

V2 was already strict about not using price momentum as a primary signal. We learned that lesson the hard way after watching too many traders chase candles. V5 keeps that discipline and adds something we wanted from the start but did not have the data architecture to support: cross-exchange confirmation and a real volume context multiplier that distinguishes a 200 BTC sweep from a 20 BTC sweep that happens to land on a quiet candle.

Here is what V5 actually scores, and why each component is in the model.

Cumulative Volume Delta is the first input, not the only one. CVD measures aggressive buy versus aggressive sell pressure on each exchange. We compute it per pair per exchange in real time, then the V5 engine asks a question that V2 could not ask cleanly. Is the CVD direction confirmed across at least two of the major venues? A long signal that fires on Binance CVD positive while Bybit and OKX CVD are flat or negative gets penalized in the score. A long signal where CVD is positive on three or more venues simultaneously gets boosted. Cross-venue confirmation is not a magic filter, but it cuts the false positives that plagued single-source signals during the March chop.

VPIN is the second input. Volume-synchronized Probability of Informed Trading is a 1996 Easley microstructure metric that measures how toxic the orderflow is. High VPIN means buy and sell volumes are extremely imbalanced over a short window, which historically precedes meaningful directional moves more often than random. V5 reads VPIN from a rolling bucket and contributes both magnitude and direction to the composite score. We do not use VPIN as a standalone trigger because it has too many regime-dependent false alarms. As a confirmation layer it earns its keep.

Order Book Imbalance is the third input. OBI measures the ratio of bid-side resting size to ask-side resting size at the top levels of the book. In V5, OBI is bucketed by depth, which means we look at the top 1, top 5, and top 20 levels separately. A pair where the top 1 OBI is bullish but the top 20 OBI is bearish is showing a thin retail bid stacked on top of a deeper institutional ask. That contradicts itself, and the score reflects that. A pair where all three depth buckets agree gets credited.

Multi-Level Order Flow Imbalance is the fourth input. OFI is the change in net resting orderbook depth over a rolling window, computed per price level. It captures the limit-order activity that book-snapshot OBI misses. Aggressive market-makers withdrawing offers count as bullish flow even before the price moves. V5 weights OFI more heavily during low-volatility regimes when book changes are more informative than tape activity, and less heavily during high-volatility regimes when most of the action is on the tape.

Whale Detection is the fifth input. We classify trades into size buckets relative to the median trade size of the last 24 hours for each pair. Trades in the top percentiles get tagged as whale activity. The V5 engine asks two questions about whale flow, not one. First, is whale net flow directional in the last fifteen minutes. Second, is the most recent whale flow consistent with the older whale flow, or is it a reversal. Persistent whale flow scores higher than a single sweep that comes after twenty minutes of opposite direction.

Funding Rate Pressure is the sixth input. Funding rates that swing far from neutral indicate crowded positioning. V5 reads funding from all five exchanges and computes a weighted z-score against the trailing seven-day distribution per venue. Extreme funding in one direction is a contrarian input. A short signal at extreme positive funding gets a confidence boost because it is fading crowd positioning, not following it.

Liquidation Squeeze Detection is the seventh input and the most directly traceable to actual entries we have logged. When liquidation maps show heavy clusters within a few percent of price and orderflow shows the tape pressing toward those clusters, the squeeze score lights up. V5 only credits the squeeze score when both conditions hold. Liquidation clusters alone do not trigger anything. The tape must already be moving in the direction of the cluster for the score to register.

Volume Context Multiplier is the eighth input and is the meta-component that ties the rest together. It is not a signal on its own but a multiplier that scales the entire composite score based on whether current volume is meaningfully above or below the trailing average. A score of 28 during a quiet 3am session is not the same as a score of 28 during peak London hours. V5 applies the volume context multiplier to all alert score thresholds, so a 28 in dead hours has to clear a higher bar to fire than a 28 during prime time.

The composite score runs from negative 100 to positive 100. Anything between negative 25 and positive 25 is treated as neutral and never fires alerts. This is the second-biggest behavioral change from V2. We stopped firing alerts when no component had real conviction, and the dedup logic now requires direction to be either buy or sell to write to the alert state cache. The neutral zone is silent.

The third behavioral change is tier-aware filtering. The Free tier on Buildix never received alerts at all, which we identified as the root cause of a retention cliff between week one and week two. We fixed the access. Trader-tier and above accounts now apply a minimum score absolute value of 25 to all signal alerts, which suppresses about 70 percent of borderline cases that the V2 engine would have surfaced. Whale tier accounts can lower that threshold to 15 if they want maximum coverage. The default protects new users from notification fatigue.

If you have been using Buildix since the V2 era, the practical difference looks like this. Fewer alerts, but the ones that fire come with cross-venue agreement, volume context, and explicit liquidation squeeze logic baked in. The signals page still shows everything, including weak setups, but the Telegram bot only pushes alerts that actually clear the composite threshold. We track signal performance publicly and in V5 the win rate on signals scoring above 35 absolute has held at 62 percent on a 14-day rolling window across BTC, ETH, SOL, and HYPE.

The engine is also versioned in our cache. The string signal_engine_version equals v5 in kv_cache, which means we can roll back instantly if we discover a regression. We have tested the V5 logic against three months of historical orderflow per pair before promoting it. Backtesting an alert engine is not a guarantee it will hold up live, which is why the rollback is one command away. So far, no regression has appeared.

Two things that V5 still does not do. It does not predict price targets. The signal is a directional confidence reading at the moment it fires, not a forecast. The trader is responsible for stop placement, position sizing, and exit logic. V5 also does not consider macro overlays like CPI prints or FOMC days. We may add a macro suppression layer in V6 if the data shows it materially improves performance during scheduled events. For now, we publish the methodology, the version string is observable, and traders can tune their thresholds inside the platform.

The biggest thing V5 changed is what it does not say. The neutral zone is the most underrated improvement. Most signal engines are loudest when they are least useful, because false confidence sells subscriptions. Silence during chop is a feature, not a bug. The market does not always have a clear direction, and the engine no longer pretends it does.

Buildix is freemium, which means anyone can open a Free tier account and watch the score live across all 530+ pairs. Telegram alerts are gated by tier, but the score itself, the component breakdown, and the signal history are visible end-to-end. The integrated AI Strategy Advisor with BYOK across OpenAI, Anthropic, Google, Groq, Mistral, and Ollama can read the V5 score and the components and explain in plain language what is driving the reading at any moment. Hyperliquid is supported from day one, which still distinguishes Buildix from most of the analytics platforms in this space.

Sign up at buildix.trade to see V5 running live. Then connect Telegram from inside the screener, set your minimum score threshold for your tier, and let the engine stay silent until something actually agrees across the components.