Predictive Intelligence
How Omnio's forecast learns your personal patterns, warns you before health dips happen, and explains why — using 18 biometric signals and Bayesian learning.
Omnio’s Predictive Intelligence engine goes beyond simple readiness scores. It learns which biometric signals predict your health changes, warns you before dips happen, and explains why — using a personalized model that gets smarter over time.
How it works
The system watches 18 biometric signals from your connected devices and learns which ones matter most for predicting your readiness. Everyone is different — for some people, HRV variability is the strongest early warning; for others, it’s sleep regularity or temperature changes.
Rather than applying the same formula to everyone, Omnio builds a personal prediction model that improves with every day of data.
The forecast
The forecast on your dashboard shows your projected readiness for the coming days, with a confidence range that reflects how certain the prediction is.
| What you need | Any wearable with daily readiness or HRV data (Oura, Garmin, Whoop) |
| Minimum data | 7 days for basic forecast; 30+ days for personalized predictions |
| Where to see it | Dashboard → Forecast card, tap for full detail |
| Updates | After each daily sync |
Adaptive horizon
The forecast extends as far as the model can predict with useful confidence. With consistent data and stable patterns, this might reach 7-10 days. During volatile periods, it may shrink to 2-3 days. The confidence cone on the chart shows exactly where predictions become too uncertain to be useful — and watching it extend over weeks is a sign your model is learning.
Confidence cone
The shaded area around the forecast line shows the range of likely outcomes. A narrow cone means the model is confident; a wide cone means more uncertainty. Day 1 predictions are tighter than day 5 — uncertainty compounds naturally.
Early warnings
When the system detects a pattern that historically precedes a readiness dip, it surfaces an early warning on your dashboard — before the dip shows up in your readiness score.
| What you need | 30+ days of multi-signal data |
| Where to see it | Dashboard → Forecast card (amber or red callout) |
| How it works | Compares current signal patterns against your personal baseline and learned predictors |
Early warnings improve over time. With less than 30 days of data, the system requires multiple signals to agree before alerting (conservative). As the model matures, it can trust single strong signals it has learned are reliable for you.
What triggers a warning
The system monitors for anomalies across all 18 signals, weighted by how predictive each signal is for your readiness. Common triggers include:
- HRV variability spike — your heart rate variability becomes more erratic than usual
- Temperature elevation — skin or core temperature deviating from your baseline
- Sleep regularity drop — inconsistent sleep-wake timing over recent days
- Autonomic withdrawal — cardiac vagal index or DFA α1 declining
- Respiratory rate increase — often an early sign of illness
Signal breakdown
Tap the forecast card to see which signals are driving the prediction. Each signal shows:
- Current value relative to your baseline (z-score)
- Trend direction — rising, falling, or stable
- Learned importance — how much this signal matters for predicting your readiness, based on your personal model
- Observation count — how much data the model has learned from
This transparency lets you understand not just what the model predicts, but why — and which aspects of your health have the most influence on your readiness.
Signals used
The model draws from these biometric signals, grouped by how they reach the system:
Core signals (from any wearable)
- Resting heart rate deviation from baseline
- HRV trend (7-day slope)
- Sleep score
- Training load (TRIMP + volume)
- Days since rest
Autonomic signals (derived automatically)
- HRV coefficient of variation — stability of your HRV, not just the average
- Cardiac vagal index — parasympathetic nervous system tone
- DFA α1 — fractal scaling of heart rate, a sensitive overtraining marker
- HRV 30-day drift — long-term trajectory invisible in day-to-day values
Sleep signals (derived automatically)
- Sleep regularity — circadian consistency over 14 days
- Sleep pressure alignment — how well your sleep timing matches circadian biology
- Sleep entropy — sleep architecture quality
Other signals
- Temperature deviation — from Oura, Whoop, or Garmin skin temperature
- Respiratory rate deviation — from overnight breathing data
- Sympathetic tone — if BLE chest strap data is available
- Autonomic resilience — HRV recovery capacity after strain
- Allostatic load — cumulative physiological stress score
Not all signals are available for all users — the model adapts to whatever data sources you have connected.
Dip patterns
When a readiness dip occurs, the system records which signals were unusual in the days leading up to it. Over time, these patterns cluster into recognizable types:
- Training overreach — elevated training load + rising HRV variability
- Sleep disruption — circadian irregularity + poor sleep architecture
- Illness onset — temperature elevation + respiratory rate increase
- Accumulated stress — rising allostatic load + declining resilience
When the early warning system detects a pattern similar to a past dip, it tells you: “This looks like the pattern before your readiness dropped on March 14.”
You can view your dip history from the forecast detail sheet under Past patterns, including which signals were anomalous and how far in advance they changed.
How the model improves
The prediction model uses Bayesian learning — it starts with population-level priors from sports science research and progressively personalizes as it accumulates your data.
| Data available | Model behavior |
|---|---|
| < 7 days | Population defaults, basic forecast |
| 7-30 days | Features begin personalizing, conservative early warnings |
| 30-90 days | Feature weights diverging from defaults, more sensitive alerts |
| 90+ days | Fully personalized, single-signal early warnings, extended forecast horizon |
The model tracks its own accuracy (mean absolute error over 30 days) and automatically widens its confidence intervals or resets to population priors if predictions drift — so it won’t confidently give you bad forecasts.