NUMBER99 LEARN
Correlation, prediction and overfitting
Historical data. Clear assumptions. Reproducible evidence.
Historical description
Frequency and gaps describe a selected dataset. They do not establish that an outcome is due, hot, cold or more likely tomorrow. Missing records and changes in source coverage can distort patterns.
Choose tests that fit the data
Two-digit result codes are categorical labels. Treating 00–99 as ordinary measurements and reporting a linear correlation may create misleading meaning. A test must match the claim and describe its assumptions.
Use unseen data
Choose the method and chronological training/test split before reviewing the holdout. Report eligible and excluded observations, baseline comparisons, sample sizes and multiple-testing adjustments. Repeatedly tuning on the holdout invalidates its status as unseen evidence.
Historical observations do not predict future outcomes.