Predixova is a production-grade, glass-box predictive analytics platform running on Google Cloud. For any symbol across four asset classes it emits a normalized 0–100 conviction score — and, critically, the entire analytical chain behind it. Clients see baseline weights, timeframe tilt, news-shock mutation, and final sub-scores, so no prediction is a black box.
Predixova is glass-box by design. This live visualizer runs the real fusion formula — select an asset class to load its baseline weight profile, then probe each channel's sub-score and watch the conviction recompute.
Pick an asset class to load its baseline weight profile, then probe the sub-scores. The fused 0–100 score recomputes in real time.
P = Σ wᵢ·sᵢ = 0.40·0.62 + 0.20·0.70 + 0.25·0.60 + 0.15·0.50 = 0.613 → 61.3
Every prediction is a weighted fusion of a Technical channel (1-D CNN embeddings matched via Vertex AI Vector Search), a News channel (per-article sentiment), a Social channel (LunarCrush Galaxy Score), and a Macro channel (CBOE VIX regime from FRED). Each is server-computed against real data and un-spoofable.
The payload exposes the full chain: baseline weights → timeframe tilt → news-shock mutation → active weights, plus each sub-score. Frontends render before/after bar charts showing exactly why the model moved — explainability is a first-class wire contract.
1.37M historical 60-bar OHLCV windows are encoded to 128-dim L2-normalized embeddings and indexed in Vertex AI. Matches are weighted by an exponential recency decay with a 1.5-year half-life, so a four-year-old pattern carries roughly 12% of yesterday's weight.
A five-tier news-impact classifier dynamically reshapes the weight vector — from zeroing noise to hard-locking news weight at 80% on systemic events. Trust boundaries ensure server-computed providers always win over client-supplied scores.
The technical channel encodes 60-bar OHLCV windows into 128-dim L2-normalized embeddings via a 1-D CNN autoencoder, indexed in Vertex AI Vector Search. Each match is weighted by an exponential recency decay with a 1.5-year half-life, so a four-year-old pattern carries roughly 12% of yesterday's weight. A five-tier news-shock classifier then mutates the weight vector — and the entire chain (baseline → timeframe tilt → shock → active weights) ships in the payload so nothing is a black box. The engine is guarded by 1,004 hermetic tests.
A four-channel predictive engine that fuses CNN pattern-matching, news sentiment, social velocity, and macro regime into an explainable 0–100 score — every weight exposed.