The NYSE Trade and Quote (TAQ) dataset, used throughout finance research and securities regulation, is generated by a Securities Information Processor (SIP), which aggregates quotes and trades from all U.S. stock exchanges at a central location. We show that the SIP systematically reports events out of sequential order: quote changes that occur after a trade are frequently reported as occurring before the trade. The source of the issue is that trades and quotes are recorded with variable latency (due to, e.g., geography) and are extremely clustered in time. The result is a look-ahead bias: the prevailing SIP quotes, ubiquitous in signing trades and measuring spreads, incorporate price impact from the trade itself. We document that the look-ahead bias leads to incorrect trade signing and downward-biased effective spreads and price impact. The errors are extreme for the large fraction of trades with high reporting latency: approximately 20% of trades are incorrectly signed, and effective spreads and price impact are understated by more than 40%. We propose a signing methodology based on exchange rules that yields 100% accuracy for a significant majority of volume and, over all trades, outperforms existing methodologies.