Statistics vs Storytelling
Why this matters
Humans are pattern machines. Markets are noisy. That mismatch produces confident nonsense: “it always dumps on Fridays,” backed by three screenshots.
This lesson is the immune system for everything else in the module. If you skip it, seasonality becomes superstition.
Biases That Fake Edges
Confirmation bias keeps confirming hits. Selection bias picks the windows that worked. Survivorship bias studies only the coins/traders that lived.
Data mining tries enough filters until something “significant” appears by chance. Overfitting memorizes noise in-sample and dies forward.
- Write hypotheses before mining charts
- Report failures, not only winners
- Prefer pre-registered rules over vibes
Sample Size & False Positives
Small n makes random streaks look like genius. Multiple comparisons (testing 100 calendar filters) guarantee false positives unless you correct for them.
Out-of-sample and forward tests exist because in-sample beauty is cheap.
- Ask n before believing a %
- Many tests → many fake winners
- Hold out data on purpose
Noise vs signal
Small samples jump between distributions; large n stabilizes.
Educational sketch — not a live signal
Practice: Edge or Random?
Train yourself to spot mined stories vs research-shaped claims. When unsure, demand mechanism, sample size, and a forward test.
Interactive · Edge or random?
Train your skepticism. Score: 0/—
BTC rose 14 of the last 16 “blue moon” Fridays.
n = 16 Fridays cherry-picked from a rare calendar filter.
Hypothesis Testing Mindset
State: “If funding > X for N intervals, next 24h return mean is ≤ 0.” Define X and N before looking. Measure. Include fees. Split regimes.
If it fails, kill it. Grief is cheaper than stubborn size.
- Define → measure → decide
- Costs are part of the hypothesis
- Ego is not a statistical method
Key Takeaways
Remember these points
- •Pattern hunger creates false edges.
- •Sample size and multiple comparisons matter.
- •Overfitting and data mining explain most calendar miracles.
- •Pre-register rules; forward-test; accept failure.
Common Mistakes
Screenshot science
Three beautiful examples are anecdotes. Count all events.
Testing until it works
Unbounded filter search manufactures significance.
Ignoring costs
A 51% directional tendency can lose money after fees and funding.
Quiz
0/3 answered1.Data mining in trading research often means:
2.A tiny sample size makes random streaks look:
3.Out-of-sample testing helps because:
Save your quiz score & progress
Create a free account to track completed lessons, save quiz results, and pick up where you left off.
Related Lessons
Price vs Time
Learn why markets move in both price and time — and why identical charts can produce different outcomes depending on when they occur.
Market Seasonality
Understand day-of-week, weekend, monthly, and yearly seasonality as drifting tendencies — not fixed laws — driven partly by institutional flow.
Discussion
Join the discussion
Sign in free to ask questions, leave notes, and learn with other traders.