papers/hanzo-analytics-ml/sections/08-evaluation.tex

\section{Evaluation}


\subsection{Prediction Performance}

\begin{table}[h]
\centering
\caption{Model Performance Metrics}
\begin{tabular}{lccccc}
\toprule
\textbf{Model} & \textbf{AUC} & \textbf{Precision} & \textbf{Recall} & \textbf{F1} \\
\midrule
Purchase (7-day) & 0.89 & 0.42 & 0.71 & 0.53 \\
Purchase (30-day) & 0.86 & 0.38 & 0.68 & 0.49 \\
Churn (90-day) & 0.84 & 0.56 & 0.73 & 0.63 \\
\bottomrule
\end{tabular}
\end{table}

\subsection{Segmentation Quality}

Segmentation quality measured by:

\begin{equation}
Silhouette = \frac{b - a}{\max(a, b)}
\end{equation}

where $a$ is mean intra-cluster distance and $b$ is mean nearest-cluster distance.

Our behavioral segmentation achieves silhouette score of 0.67, compared to 0.41 for demographic-only segmentation.

\subsection{Business Impact}

\begin{table}[h]
\centering
\caption{Business Outcomes}
\begin{tabular}{lcc}
\toprule
\textbf{Metric} & \textbf{Before} & \textbf{After} \\
\midrule
Campaign ROI & 1.8x & 4.1x \\
Churn Rate (90-day) & 23\% & 17\% \\
Reactivation Rate & 8\% & 19\% \\
LTV Accuracy & $\pm$32\% & $\pm$14\% \\
\bottomrule
\end{tabular}
\end{table}

\subsection{Latency Performance}

\begin{table}[h]
\centering
\caption{Prediction Latency}
\begin{tabular}{lcc}
\toprule
\textbf{Percentile} & \textbf{Feature Fetch} & \textbf{Prediction} \\
\midrule
P50 & 8ms & 3ms \\
P95 & 23ms & 8ms \\
P99 & 45ms & 15ms \\
\bottomrule
\end{tabular}
\end{table}