Evaluation & Results: Can Machine Learning Beat Lottery Randomness?

This section presents the results of a structured evaluation designed to understand how machine learning performs in a system designed to be random.

Using the Kerala State Lottery as a controlled environment, multiple algorithmic approaches were tested under real-world conditions. The focus is on measuring predictive performance and evaluating the practical relevance of any detected signal.

Experimental Setup

• Evaluation period: 91 days (out-of-sample)
• Total observations: 33,000+
• Algorithms tested: 36 (6 families × 6 time windows)
• Evaluation method: Book-level simulation
• Ticket price: ₹50
• Statistical testing: One-sided binomial with Bonferroni correction

This setup ensures results reflect real-world performance without data leakage or overfitting.

Performance Metrics

Hit Rate

Measures how often predicted numbers match actual outcomes. Baseline (random): 10%

Lift

Represents improvement over random selection. Example: 1.15 = 15% improvement

Statistical Significance

Evaluates whether improvements are likely due to chance. Adjusted for multiple comparisons.

Return on Investment (ROI)

Simulates real-world financial outcomes to assess practical impact.

Results Overview

• Best hit rate: 11.55%
• Random baseline: 10%
• Lift: 1.154
• Statistical significance: p < 10⁻⁶

The models demonstrate measurable improvement over random selection, indicating the presence of detectable statistical signals.

Model Performance Insights

Positional Digit Analysis

• Consistent performance across time windows
• Suggests minor structural variation in digit distribution

Composite Scoring

• Highest peak performance
• Combines multiple signals effectively

Short-Term Models (15–30 days)

• Stronger performance than long-term models
• Indicates signals may be time-sensitive

Recency-Based Models

• Based on recent occurrence patterns
• Did not show consistent improvement

Statistical Significance

• 17 out of 36 models achieved statistically significant results
• All positional digit models were significant
• Several composite models passed strict thresholds

This confirms the presence of weak but measurable signals within the system.

ROI Analysis

• ROI range: −54% to −85%

Example:

• Random expected return: ~₹20 per ₹50 ticket
• Best model return: ~₹23 per ₹50 ticket

Lottery participation is not structured as an investment approach aimed at generating consistent returns like trading or financial assets. It is more appropriately viewed as an entertainment-based activity, where outcomes are governed by chance rather than predictable return patterns.

From a practical perspective, participation is best approached with moderation, using only discretionary or excess funds allocated for entertainment purposes. This study does not promote or encourage gambling behavior, but rather provides a structured evaluation of machine learning performance within such systems.

Understanding Practical Constraints

Prize Distribution

High-value prizes are rare, while most wins are lower-value.

Execution Constraints

Practical limitations affect how strategies can be applied in real-world scenarios.

System Structure

Lottery systems are designed with predefined payout structures that influence outcomes.

Key Insight: Detectable vs Practical Signal

• Signals can be identified and measured
• Their real-world impact depends on multiple external factors

Final Takeaway

• Machine learning can outperform random selection under controlled evaluation
• Performance improvements are measurable and statistically valid
• Real-world outcomes depend on system-level constraints