The story ·
Which habits, moods and circumstances actually move exam scores? A scoring engine tested every pair, triple and quadruple of factors in 500,000 computer-generated student records.
Study hours, stress, internet, sleep, family, motivation… which of them really connect to how students do, and which only look like they do?
A scoring engine tested every pair, triple and quadruple of factors in 500,000 student records. It kept only connections that were statistically significant after correcting for thousands of tests and that reappeared in a held-out half of the data. Some are obvious. Some are curves, synergies and hidden pockets that no correlation table would show.
These students are not real. The data was generated by a computer program, and its author deliberately planted the effects you’ll find. Treat this as a demonstration of how to discover connections, not as evidence about real students.
Data. 500k Student Performance and Behavior Dataset by Waddah Ali on Kaggle, licensed CC BY-NC-SA 4.0. According to its author it is fully synthetic: generated with NumPy (seed 42) with deliberately embedded effects (an inverted-U for stress, an internet × study-hours interaction, library visits derived from study hours, and two injected outlier groups). Demographic and institutional variables were generated without any link to outcomes, so they sit in the “no signal” ring.
Scoring engine. On 200k rows split into discovery and holdout halves, the engine searched:
All p-values are Benjamini–Hochberg corrected, and every finding must replicate on the holdout half.
Score = effect × significance × replication × non-obviousness (0–100). Obvious, mechanical or explained-away links are down-weighted; curves, synergies and pockets are up-weighted.
Effect sizes. r = Pearson correlation; η = correlation ratio (captures curves); Cohen’s d = standardized gap between a pocket and everyone else; ΔR² = extra variance explained by a combination beyond separate effects; Δr = change in correlation once a third factor is held fixed.
Live mode. The population filters recompute associations in your browser on a 30,000-student sample, so edge strengths respond as you narrow the group.
This visualization is a derivative work and is shared under CC BY-NC-SA 4.0. Non-commercial use only. Associations are not causes.