TradingAgents 0.4.0 fixes five look-ahead leaks in backtests

TradingAgents closed five separate ways its own backtests were cheating. The v0.4.0 release pins FRED macro data requests to the as-of vintage, so a backtest can't see a GDP number that got revised weeks after the fact. StockTwits and Reddit sentiment now gets trimmed to the actual analysis window instead of pulling in posts from after the decision point. Decision-log memory records when each outcome became knowable, so the system stops feeding itself lessons about trades that hadn't resolved yet on the date it's supposedly deciding.

Two smaller fixes round it out. An unparseable Portfolio Manager rating used to default to a tradeable Hold, turning a parsing failure into a live position with no error surfaced anywhere. It now throws a REVIEW sentinel instead. Debate openers stop inventing an opponent's argument when no argument was made.

Each of these is a look-ahead leak, the class of bug where a backtest sees information it wouldn't have had in real time. They're the reason paper trading systems post great historical returns and then underperform live. A framework this popular, with 101,945 GitHub stars, had five of them stacked at once: macro data, two sentiment sources, memory retrieval, and a rating parser all leaking future information into past decisions.

If you're running TradingAgents or forked it, your historical backtest numbers from before this release don't mean what you thought. The FRED and sentiment fixes alone change what data was visible at each decision point, so any Sharpe ratio or win rate calculated pre-0.4.0 needs a rerun. The release notes don't say how much the fixes move the numbers, only that the leaks existed.