Drishti
What Changes
What is stationary vs non-stationary?
What Changes?
TL;DR
- Separate stationary structure from drifting signals — model each differently.
- Every system has multiple clocks — fast noise, slow trends, rare regime shifts.
- Non-stationarity is the default — assuming constancy is the risky special case.
How to use this lens
- Inventory what you treated as fixed in your mental model.
- Classify rates of change — milliseconds, days, years, “never in practice.”
- Mark non-stationary drivers — seasonality, growth, policy, fashion, hardware generations.
- Ask what would break your model if a parameter drifted 10%.
- Identify stable coordinates — things that change slowly enough to plan against.
- Note feedback — does the system change because you measured it?
Core concepts
Takeaway: Stationarity is a local approximation.
Queueing formulas assume arrival rates hold still for a while. That is often useful for the next hour, never for the next decade. Streaming analytics exists because the world refuses to sit still — windows slide because history stops predicting.
Takeaway: Multiple timescales stack.
Sleep cycles change minute to minute; sleep need shifts across a lifetime. Offices rewrite policies quarterly while Slack messages churn by the second. Good analysis ** picks a timescale** before picking a tool.
Takeaway: Regime changes are discrete jumps.
A startup before and after product-market fit is not the same system with tweaked parameters — incentives, bottlenecks, and failure modes reorganize. Treat regime shifts as new existence questions, not noise.
Change taxonomy
| Type | Signal | Risk if ignored |
|---|---|---|
| Noise | Random fluctuation | Overfitting |
| Trend | Directional drift | Stale forecasts |
| Seasonality | Repeating calendar | Wrong baseline |
| Regime shift | Rule change | Catastrophic model error |
Mini-examples
- Sleep: circadian phase rotates daily; sleep pressure accumulates hourly; aging shifts baseline across years.
- Office: headcount grows; strategy pivots; tooling generations replace every few years.
- Blood: heart rate varies beat-to-beat; fitness changes capacity over months.
Curriculum bridge
When change becomes formal:
- Eventual Consistency — systems that accept temporary mismatch while converging
- Streaming Analytics — computation over evolving streams
Try it now
Pick one metric you watch (steps, error rate, mood). Is it noisy, trending, or seasonal? What would a chart at a different timescale reveal?
Going deeper
Takeaway: Observability changes behavior.
People sleep differently when tracked. Teams game metrics once bonuses attach. The act of measurement is part of the dynamics — not an external camera.
Takeaway: Slow variables trap fast optimizers.
Optimizing this week’s throughput can erode next year’s reliability. Change lens pairs with what persists — find variables that move slowly and anchor plans there.
Takeaway: Adaptation is change you want.
Learning systems should change. Distinguish destructive drift (model rot) from productive adaptation (updated weights). The difference is whether performance on held-out goals improves.
If you skip this lens, you will confidently apply a formula whose assumptions expired yesterday.