cs.LGNov 16, 2026

Reading or Remembering? Measuring Temporal Contamination in Language Models

Authors: Arunkumar V, Gangadharan G. R., S. Senthilkumar

Organizations: UCE BIT Campus, Anna University, Tiruchirappalli, India · NIT Tiruchirappalli, Tiruchirappalli, India

Abstract

Language models are now used to read central-bank text for economic forecasting, but most were trained on text written after the documents they read. A model reading a 2021 FOMC statement may already know that inflation rose in 2022, so part of its apparent skill can be memory of the outcome, a temporal form of test-set contamination. We measure this memory on FOMC statements and minutes from 2000 to 2026 by comparing modern open models with chronologically consistent models, which are trained only on text from before a fixed date. Asked about past GDP outcomes given only a date, models that cannot know the answer still reach an AUC of up to 0.64 from priors on dates and eras, so chance is the wrong baseline. Against these memory-free models on the same questions, most modern models recall which years had strong growth. Membership-inference scores sort documents by era and document type rather than by whether a model saw them, as exact counts in an open pretraining corpus show. In real-time nowcasts, every lookahead premium we estimate on US GDP is smaller than the forecast comparison can detect at macroeconomic sample sizes, and none on GDP or inflation survives correction for multiple testing. Lookahead audits should therefore measure memory directly against memory-free models on the same items.

Figures & tables

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Mar 12, 2026

DatedGPT: Preventing Lookahead Bias in Large Language Models with Time-Aware Pretraining

Large language models pretrained on internet-scale data risk lookahead bias in forecasting tasks, as they may have already seen the true outcome during training. To address this, we present DatedGPT, a family of twelve 1.3B-parameter language models trained from scratch on approximately 100 billion tokens each with strict annual data cutoffs spanning 2013 to 2024, together with DatedInstruct, an instruction dataset grounded in each year's documents to prevent leakage during post-training. The models are competitive with open models of similar scale, and perplexity-based probing confirms that each model's knowledge is bounded by its cutoff year. On stock return prediction over 61,000 firm-day news headlines, DatedGPT-instruct achieves an annualised Sharpe ratio of 3.203.20 under the lookahead-bias-free setup. Lookahead-biased models, whose training data covers the outcome period, add a lookahead premium of 26.426.4 b.p. per standard deviation, significant at the 1% level. The series thus enables direct analysis of lookahead bias in financial forecasting. We provide an interactive web demo that allows users to query and compare responses from models across different cutoff years, available at www.datedgpt.com.
Apr 24, 2026

Scaling Point-in-Time Language Models: Enabling Economic Evaluation of Embeddings

Large language models trained on unrestricted internet corpora inevitably embed information from the future, introducing lookahead bias that compromises the validity of backtests and causal inference in finance and the social sciences. Point-in-time language models--trained exclusively on text available up to each calendar date--eliminate this leakage by construction, but existing efforts typically produce models that lag substantially behind their unconstrained counterparts. We show that this performance gap can be substantially narrowed through scale. Training decoder-only transformers with up to 4 billion parameters on 1 trillion chronologically filtered tokens from FineWeb, we construct a sequence of monthly model checkpoints spanning 2013-2024. Across a range of common-sense reasoning and language understanding benchmarks, our models approach the performance of leading open-weight models of comparable size (e.g., Gemma-3-4B and LLaMA-7B) trained on temporally unrestricted data, although a performance gap remains on several tasks. Instruction fine-tuning via LoRA further improves downstream usability. We release the complete pipeline--including dataset construction, training infrastructure, and evaluation code--to enable reproducible point-in-time language modeling and to support research applications that require strict temporal validity.
Mar 26, 2026

MemGuard-Alpha: Limits of Membership Inference for Detecting and Filtering Memorization-Contaminated Signals in LLM-Based Financial Forecasting

Large language models are increasingly used to generate financial alpha signals, but many have memorized the historical data in their training corpora, producing apparent accuracy that collapses out of sample. Membership inference attacks (MIA) have been proposed as a diagnostic. What has not been established is whether MIA scores are informative about memorization in this setting, or whether signal-level filtering built on them helps once realistic costs are applied.We introduce MemGuard-Alpha, comprising a composite contamination score combining five MIA methods with a temporal proximity feature, and Cross-Model Memorization Disagreement, which exploits variation in training cutoffs across models. We then audit both. Across seven LLMs (124M-7B), 50 S&P 100 constituents, 42,800 prompts and 299,600 prompt-model MIA scores spanning 2019-2024, three findings emerge.First, where in-sample status is defined by a training cutoff, a temporal proximity feature recovers that label perfectly (ROC-AUC 1.000) because it is a monotone transform of the defining variable; any composite score containing such a feature reports separation that is arithmetic rather than detection. Second, the discriminative power of the MIA scores is largely attributable to scale differences between models: asked at a fixed date which models had that date in training, raw scores appear highly informative (AUC up to 0.99), but under three independent within-model normalizations discrimination falls to 0.487-0.537. Third, with transaction costs applied symmetrically, no filtering variant improves risk-adjusted performance over the unfiltered ensemble, none attains significant Fama-French five-factor alpha, and excluding the single weakest model outperforms every contamination-based filter.We report these as negative results with the failure modes that produced them, and release all artifacts needed to reproduce them.