"This section uses a Gaussian copula to model joint default risk," she said. "But, the underlying data set includes high-yield bonds from the 2008 crisis. The Gaussian copula famously failed during that period because it underestimates tail dependence. Your model is using a tool that already broke once in exactly the conditions it's supposed to predict." She didn't wait for a response. She wrote the correction, a Clayton copula with adjusted tail parameters, and showed how the risk profile changed.
The difference was significant. Under Harrington's model, the probability of a simultaneous default event was 0.3%. Under Whitney's correction, it was 4.1%. Thirteen times higher. Callaway stood up again. He didn't walk to the board this time. He didn't need to. He turned to the investor beside him and said, clearly, "I'm pulling my allocation first thing Monday morning." The words spread through the room like fire through paper. Two more investors reached for their phones.
One woman closed her portfolio binder and slid it into her bag. Harrington felt it. His jaw tightened. He stood up and walked to Whitney. "You're cherry-picking," he said. "You're finding minor issues and blowing them up to look smart. The core model is sound." Whitney didn't argue. She turned back to the projection and moved to page six. Then Harrington played dirty. He walked to Slade and whispered something.
Slade hesitated, then typed on his laptop. The projection flickered. When it came back, the layout had changed. Pages had been reordered, and two sections Whitney had already reviewed were swapped with new ones she hadn't seen. The parameters looked different. The notation had shifted. They were changing the model while she was auditing it. Whitney stared at the screen. Her marker hand paused. For the first time all evening, her rhythm broke.
She blinked. Looked at the whiteboard where her previous corrections sat. Looked back at the screen. The numbers didn't match anymore. A murmur went through the crowd. Harrington smiled, wide, full of teeth. "Problem?" he said. "I thought you were a genius." Whitney closed her eyes. Three seconds. Four. When she opened them, she started from the top of the rearranged page. She didn't complain. She didn't call it unfair.
She rebuilt her frame of reference from scratch, tracing the new layout back to the original architecture she'd already mapped in her head. It cost her six minutes, six minutes she did not have. Pages 6 and 7 restructured, clean. Whitney noted them as structurally sound and moved on. She wasn't performing. She wasn't showboating. She was auditing. Methodically. Relentlessly. Without emotion. That made it worse for Harrington. If she had been arrogant, he could have attacked her character.
If she had been wrong, he could have attacked her math. But she was quiet. And she was right. And he had no weapon for that. Page 8. Whitney found the third significant error. A bootstrap resampling procedure that used overlapping windows, introducing autocorrelation into the confidence intervals. She corrected it. The model's claimed 94% accuracy dropped to 71%. Harrington's face went white. Slade leaned over and whispered, "Grant, we should stop this. Now."