How Stack Almanac's Correlation Engine Finds What Works for You

Most supplement advice is based on population-level research: studies that found an average benefit across hundreds or thousands of participants. But averages hide individual variation. A supplement that works for the "average" person might do nothing for you, or it might work even better.

Stack Almanac's correlation engine solves this by learning from your data. Here's how.

The basic principle

The correlation engine asks a simple question: on days you take a specific supplement, do your outcomes differ from days you skip it?

For every active supplement in your stack, it compares:

• Days taken (confirmed or assumed) vs days skipped
• Across six outcome metrics: energy, mood, sleep quality, focus, stress, and overall well-being
• Plus health metrics: sleep hours, heart rate variability (HRV), and resting heart rate

When enough data accumulates (at least 3 days taken and 3 days skipped) the engine calculates the difference. If your average sleep quality is 3.8 on magnesium days and 3.1 on non-magnesium days, that's a meaningful signal.

Three-state accuracy

Not all data is equally reliable. Stack Almanac uses a three-state logging model:

• Confirmed (you tapped "taken"): weighted at 100%
• Assumed (auto-marked after your window passed): weighted at 50%
• Skipped (you explicitly skipped): weighted at 100%

This weighting prevents false correlations. Assumed days contribute less to the analysis because they're less certain. The more you confirm or explicitly skip, the more accurate your correlations become.

The advisor acknowledges this data quality in its responses: "Based on 12 confirmed and 4 assumed magnesium days..."

Combination discovery

Single-supplement correlations are useful, but some supplements work better together. The correlation engine also analyses supplement pairs:

It compares four states:

1. Both taken: supplement A and B on the same day
2. A only: A taken, B skipped
3. B only: B taken, A skipped
4. Neither: both skipped

This reveals synergistic combinations from your personal data. For example, you might find that magnesium glycinate alone improves your sleep by 15%, but magnesium plus L-theanine together improve it by 30%. That combination effect wouldn't show up in single-supplement analysis.

Cycle-phase correlations

For users who track menstrual cycles, the engine runs a separate analysis by cycle phase. This reveals whether a supplement works differently during the follicular phase (first half) versus the luteal phase (second half).

This is genuinely novel. Most supplement research doesn't stratify by menstrual phase, yet hormonal fluctuations meaningfully affect supplement response.

What it takes to get useful data

The correlation engine needs time and variation:

• Minimum 2-3 weeks of consistent tracking before patterns emerge
• Both taken and skipped days: if you never skip a supplement, there's no comparison point. This is why trials (temporary removal) are valuable.
• Consistent outcome logging: rating your energy and sleep daily, even briefly
• Patience: biological changes are gradual. A supplement that takes 2 weeks to build up (like creatine or omega-3) won't show immediate correlations.

How it feeds the advisor

The correlation engine's findings are automatically injected into the AI advisor's context. When you ask "is my magnesium working?", the advisor doesn't just cite research. It cites your data:

"Based on your last 30 days: on the 18 days you took magnesium glycinate, your sleep quality averaged 4.1/5. On the 7 days you skipped, sleep averaged 3.3/5. That's a 24% improvement, which is significant."

This closed loop (take, track, correlate, advise) is what makes Stack Almanac's recommendations genuinely personalised rather than generic.

The limitations

Correlation isn't causation. The engine identifies patterns, not proof. Confounding variables exist: maybe you sleep better on magnesium days because you also exercise more on those days. The advisor explicitly frames findings as correlations and recommends controlled trials (isolating one variable at a time) when the data is ambiguous.

The engine requires a minimum difference threshold (0.3 on a 5-point scale, or 10% change) to flag something as significant. Small effects below this threshold are filtered out to reduce noise.

Why this matters

The supplement industry is built on population-level evidence and influencer anecdotes. Neither tells you what works for your specific body. The correlation engine gives you a third source of truth: your own data, analysed systematically over time.

It won't replace clinical research or blood testing. But it adds a personal layer of evidence that neither of those can provide, because they don't know what you took yesterday, how you slept last night, or how that compares to the pattern over the last month.