| # HyperFlow: ML Benchmark & Evaluation Methodology |
| **System Evaluation Criteria & Mathematical Proofs** |
|
|
| --- |
|
|
| ## 1. Monte Carlo Simulation Framework |
| To evaluate model stability under highly noisy and censored hyperlocal conditions, HyperFlow runs a **500-trial Monte Carlo simulation framework**. Each trial generates simulated order logs, telemetry GPS vectors, user complaints, and cancellation coordinates. |
|
|
| ``` |
| +---------------------------------------------+ |
| | Monte Carlo Data Generation (500 Trials) | |
| +---------------------------------------------+ |
| | |
| +----------------------+----------------------+ |
| | | |
| [Out-of-Stock Censoring] [Storm Surge Telemetry] |
| | | |
| [OLS vs Tobit Estimator] [EWMA vs Gated Forest] |
| | | |
| +----------------------+----------------------+------+ |
| | |
| +------------------------------+ |
| | Wasserstein Distance Calc | |
| +------------------------------+ |
| ``` |
|
|
| --- |
|
|
| ## 2. Tobit Parameter Recovery & Wasserstein Distance |
|
|
| ### Tobit MLE Log-Likelihood |
| Our Tobit regression model accounts for censored sales values on stockout days. The parameter recovery process estimates how close our estimated coefficients ($\hat{\beta}$) are to the true latent demand coefficients ($\beta_{\text{true}}$). The log-likelihood function minimized by the backend L-BFGS-B solver is: |
| |
| \[L(\beta, \sigma) = \sum_{y_i > 0} \ln \left( \frac{1}{\sigma_i} \phi\left(\frac{y_i - X_i\beta}{\sigma_i}\right) \right) + \sum_{y_i = 0} \ln \left( 1 - \Phi\left(\frac{X_i\beta}{\sigma_i}\right) \right)\] |
| |
| Where: |
| * \(\phi(\cdot)\) is the standard normal PDF. |
| * \(\Phi(\cdot)\) is the standard normal CDF. |
| * \(\sigma_i\) is the heteroscedastic scale parameter modeled as \(\log(\sigma_i) = X_i\gamma\). |
|
|
| ### Wasserstein Distance Evaluation |
| To prove that our Tobit model recovers the true demand distribution rather than simply predicting the mean, we calculate the **1st Wasserstein Distance** (Earth Mover's Distance) between the predicted demand distribution (\(P_{\text{pred}}\)) and the true demand distribution (\(P_{\text{true}}\)): |
|
|
| \[W_1(P_{\text{pred}}, P_{\text{true}}) = \int_{-\infty}^{\infty} |F_{\text{pred}}(x) - F_{\text{true}}(x)| \, dx\] |
|
|
| * **OLS Baseline**: As censoring rates increase to 60%, the OLS Wasserstein distance degrades to **12.38** due to severe downward bias. |
| * **Tobit MLE**: Our model maintains a Wasserstein distance of **3.65**, preserving the underlying distribution alignment. |
|
|
| --- |
|
|
| ## 3. Gated Random Forest ETA Convergence |
| Display ETA jitter is suppressed by classifying location updates. The Gated Classifier evaluates the residual convergence rate over a moving window: |
|
|
| \[\text{Residual}(t) = |\text{ActualDeliveryTime} - \text{PredictedETA}(t)|\] |
|
|
| If the location update fails to converge toward the actual delivery trajectory, it is flagged as a transient GPS jump: |
|
|
| \[\text{Jitter Flag} = \text{Classifier}\left( \frac{v_{\text{rider}}}{v_{\text{zone}}}, \text{GPS\_Accuracy\_Index}, \Delta\text{Heading} \right)\] |
|
|
| * **Jitter Active**: The smoothing parameter $\alpha$ drops to **0.15** to hold the consumer's display clock stable. |
| * **Real Delay Active**: The smoothing parameter $\alpha$ rises to **0.70** to pass the delay information immediately. |
|
|
| --- |
|
|
| ## 4. Anti-Arbitrage Proximity Thresholds |
| The Cancelled Order Resale (CORO) engine restricts discount exploits by verifying buyer proximity coordinates using the **Haversine Distance Formula**: |
|
|
| \[d = 2R \arcsin \left( \sqrt{\sin^2\left(\frac{\Delta\phi}{2}\right) + \cos(\phi_1)\cos(\phi_2)\sin^2\left(\frac{\Delta\lambda}{2}\right)} \right)\] |
|
|
| Where: |
| * \(\phi\) is latitude, \(\lambda\) is longitude. |
| * \(R\) is Earth's radius (6371 km). |
| * **Exclusion Policy**: Any claim where \(d < 15\text{ meters}\) from the original canceller's coordinates triggers a co-location exclusion alert. |
|
|
