Agentic Quantitative Trading: Why Backtest Performance Often Fails in Live Markets

I read it looking for edge. What I found instead was a confirmation of what every systematic trader already suspects — strong backtest forecasting does not reliably translate into live-market performance.

The pipeline reality check

The authors map the modern agentic quant pipeline into five working modules. That is a useful taxonomy if you have ever tried to bolt an LLM onto a research stack and wondered where the latency actually lives. The blunt finding: the bulk of deployed systems still concentrate on signal discovery. Factor mining gets the conference papers. Order execution and risk management get the afterthought slides.

That distribution tells you where the alpha theater sits. Signal generation is the photogenic half of the workflow — it produces charts, it produces metrics, it produces papers. Execution and risk are the plumbing. Nobody publishes a paper on slippage calibration, and that is precisely why the plumbing is where the edge lives.

Where the edge leaks

The survey's core warning is one I hammer in this column constantly: a model that looks brilliant in-sample is not a model with edge. Forecasting capability — high information coefficient, decent Sharpe on the backtest — is a necessary but insufficient condition. The jump from backtest PnL to realized PnL is where most agentic stacks hemorrhage capital. Transaction costs, regime drift, queue position, and crowding collapse the assumption set the moment you go live.

If you are deploying any agentic research output right now, the question is not "does it predict." The question is "does it survive the friction budget, the capacity constraint, and the drawdown tail." The arXiv work is essentially a long-form admission that the industry has not solved the second question. Forecasting is solved on slides. Forecasting-to-execution is unsolved in production.

What I am watching

Three takeaways from the survey worth tracking for systematic traders running agentic components.

First, evaluation methodology. Look for out-of-sample frameworks that explicitly price in slippage, latency, and capacity. If a paper or product does not show those numbers, treat the edge as imaginary until proven otherwise.

Second, module coverage. Any vendor pitching an "AI quant" that only describes the signal layer is selling you the photogenic half. Ask what they do for execution and risk, or assume they do nothing.

Third, sample size on live deployment. Backtest years are cheap. Live trading months are data. Demand the latter before you allocate capital.

The survey is a snapshot of where the field is, not where it will be. Right now the field is still doing what retail traders do with indicator scripts: chasing the prettiest signal, ignoring the plumbing. The edge has always lived in the plumbing. A new architecture does not change that arithmetic.