Skip to content

Temporal Knowledge Leakage in LLM Backtesting (Model-Level Future Information Contamination) #805

Description

@KenCheung-AIxFinance

First of all, excellent work on TradingAgents.
The architecture, agent coordination design, and historical data pipeline are very impressive.

However, I would like to raise a concern regarding a potential evaluation issue in LLM-based backtesting:

Temporal Knowledge Leakage from the Base LLM

From the documentation and implementation, TradingAgents already appears to handle traditional future leakage correctly:

  • date-aware data fetching
  • historical news restriction
  • timestamp-based retrieval
  • historical OHLCV snapshots

This prevents explicit future data exposure.

However, there is another layer of leakage that may still exist:

Model-Level Future Knowledge Contamination

For example:

  • using a 2026-trained LLM
  • while backtesting decisions in 2023

Even if the prompt context is historically constrained, the model weights themselves may already encode:

  • macro outcomes
  • company trajectories
  • major crises
  • sector bubbles
  • future market narratives

Examples:

  • NVDA AI boom
  • FTX collapse
  • SVB failure
  • post-rate-hike market behavior
  • inflation outcomes

This means the model may unintentionally perform:

  • hindsight reasoning
  • retroactive narrative fitting
  • future-aware macro interpretation

rather than genuine ex-ante reasoning.


Question

Has the team considered this issue formally?

Specifically:

  1. Do you view TradingAgents backtests as:

    • historically grounded simulations
    • or strictly causal backtests?
  2. Have you explored:

    • time-frozen checkpoints
    • older foundation models
    • cutoff-constrained continual pretraining
    • temporal contamination evaluation
  3. Are there plans to introduce:

    • model cutoff declarations
    • leakage benchmarks
    • temporal purity evaluation protocols

Possible Research Direction

A potentially rigorous protocol could be:

Backtest Period Model Knowledge Cutoff
2023 <= 2022
2024 <= 2023

with:

  • retrieval cutoff
  • embedding cutoff
  • reranker cutoff
  • memory cutoff

all aligned temporally.

Beside above questions, what solutions has your team come up with for this problem?


I believe this is becoming an important issue for the entire LLM-for-finance research community, especially as reasoning models become stronger.

Would love to hear the team's perspective on this.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions