Separating ML Trading Signals from Portfolio Risk Controls

# How are risk management practices applied to ML/AI-based automated trading systems


# How are risk management practices applied to ML/AI-based automated trading systems












A potential issue with automated trading systems, that are based on Machine Learning (ML) and/or Artificial Intelligence (AI), is the difficulty of assessing the risk of a trade. An ML/AI algorithm may analyze thousands of parameters in order to come up with a trading decision and applying standard risk management practices might interfere with the algorithms by overriding the algorithm's decision.

What are some basic methodologies for applying risk management to ML/AI-based automated trading systems without hampering the decision of the underlying algorithm(s)?

Update: An example system would be: Genetic Programming algorithm that produces trading agents. The most profitable agent in the population is used to produce a short/long signal (usually without a confidence interval).

## Answer by Shane (score 13, accepted)

https://quant.stackexchange.com/a/53

The risks involved in trading is everywhere and always a multifaceted thing: it includes the volatility of the selected asset, the leverage and concentration of the porfolio, whether there is a stop loss, a hedge, etc. Also, risk management is frequently not tied to the "alpha model" directly (e.g. VaR, shortfall, and scenario testing).

For instance, one well known way of sizing a position is the Kelly formula:

$f^{*} = \frac{bp - q}{b}$

This makes no assumptions about the directional model that is used to enter the position. You can infer the values (e.g. probability of winning) from a historical simulation, regardless of whether the model is black-box, grey-box, or white-box.

## Answer by shabbychef (score 20)

https://quant.stackexchange.com/a/68

ML/AI systems are susceptible to a number of risks not traditionally discussed in risk management:

- What I call 'backtest arbitrage'. In the process of automated model generation and testing, your machine learner may discover, exploit, and concentrate on irregularities in your backtesting system which do not exist in the real world. If, for example, your fill simulation is erroneous, you have not accounted for borrow costs, forgot to deal with dividends properly, etc., sufficiently powerful search techniques will find strategies which capture these nonexistent 'arbs'.

- If you sequentially generate, test, and refine many trading models, you run into the problem of 'datamining bias'. Here one has used the same data to simultaneously select the best model and estimate its performance via backtest. The estimate will be positively biased, and the size of the bias can be difficult to estimate if one has not kept careful track of all the strategies tested.

- Blackbox models are often subject to non-stationarities of the 'Grue and Bleen variety'. That is they may behave radically different out of sample due to non-stationarities of their input data and discontinuities in their processing of input data. An example would be an AI strategy which first checks if VIX is above 60, then trades one substrategy, otherwise it trades a different one. Your backtest period may contain little data in the 'over 60' regime, and you may find yourself in such a regime.

Regrettably many of these issues exist at the human level, and there is little one can do statistically to detect them or correct for them. They require great attention to process.

## Answer by Dan (score 3)

https://quant.stackexchange.com/a/2444

It depends on what the strategy does.

For a long/short signal on an equity symbol, one way is to look at the options prices / implied volatility for that symbol. Your system should give an expected timeframe and profitability, so the risk involved could be quantified by the price of buying options to insure yourself against losses compared to your expected returns.

For a more complicated symbol, you can attempt to approximate using a basket of options.

For a short-term microstructure trade (which is IMHO the arena in which ML/AI-type strategies are most useful, mainly because proper quant analysis often has little to say), there is very little you can in terms of principled risk estimates. Rather, you must rely on simulation, and backtesting. I especially recommend adding in simulations for totally disastrous fictional scenarios. For this kind of trade, any estimation is rough at best, so use a healthy dose of pessimism and superstition.

## Answer by lehalle (score 2)

https://quant.stackexchange.com/a/3281

The risk is not linked to the decision process but to your inventory, independently from the signals that triggered the buys and sells: you can monitor the inventory as usual.

If you are talking of taking into account the fact that you change your inventory more often because you use computer-based signals, it is more complex. You need to control the dynamics of your trading algorithm to be sure that it will not take positions in one millisecond that will dramatically increase your risk.

## Answer by Russlan Ramdowar (score 0)

https://quant.stackexchange.com/a/85768

I would separate the alpha model from the risk controller.

The learning algorithm should estimate expected returns, probabilities or relative scores. A separate layer should translate those forecasts into feasible positions. One possible formulation is

$ x_t=\arg\max_x \left( \hat{\mu}_t^\top x -\lambda x^\top\hat{\Sigma}_t x -C_t(x-x_{t-1}) \right) $

subject to limits on gross and net exposure, concentration, factor exposure, liquidity, leverage and loss.

This does not require the risk layer to understand every internal feature used by the ML model. It requires the risk layer to understand the proposed exposure and its consequences.

For the genetic-programming example in the question, I would use four distinct layers.

#### 1. Model-selection risk

Selecting the most profitable agent from a large population is an extreme-value selection problem. Its backtested performance is biased upward even if every individual backtest is calculated correctly.

I would preserve the results of the entire population and every parameter trial—not only the winning agent—and evaluate the complete research process using genuinely out-of-sample data. The Probability of Backtest Overfitting framework is directly relevant here:

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2326253

The winning agent should not be allowed to select itself and then use the same observations to estimate its own risk.

#### 2. Forecast uncertainty

If the agent produces only a long/short signal, the surrounding research process should estimate uncertainty.

Possible approaches include block bootstrapping, performance across walk-forward folds, dispersion across independently trained agents, and conditional results across volatility or liquidity regimes.

When uncertainty or model disagreement is unusually high, the system can reduce exposure or abstain. Abstention is often more defensible than forcing every signal into a full position.

#### 3. Portfolio and execution risk

I would apply controls independently of the model's preferred direction:

- gross and net exposure limits;

- single-name, sector and factor concentration limits;

- volatility or expected-shortfall budgets;

- liquidity and participation-rate constraints;

- borrow availability and financing-cost checks;

- turnover, spread and market-impact estimates;

- stale-data and abnormal-input detection;

- daily-loss, drawdown and kill-switch rules;

- stress tests for gaps, volatility shocks and correlation convergence.

The algorithm proposes an exposure. The portfolio controller determines whether that exposure fits within the strategy's risk budget.

#### 4. Model and operational governance

Every production decision should be reproducible from a dated data snapshot, model version and configuration. I would also require independent validation, champion/challenger comparisons, drift monitoring and explicit conditions for suspension or retraining.

The NIST AI Risk Management Framework provides a useful general lifecycle structure:

https://doi.org/10.6028/NIST.AI.100-1

The Federal Reserve's revised 2026 model-risk guidance is also a useful governance reference, although its legal applicability depends on the institution:

https://www.federalreserve.gov/supervisionreg/srletters/SR2602.htm

I would be particularly cautious with Kelly sizing in this example. A probability of winning estimated from the same optimized backtest is not a clean input to Kelly. It contains selection bias and model uncertainty. If Kelly is used at all, I would use independently validated probabilities and a fractional allocation.

Therefore, risk management does "override" some model decisions—but that is its purpose. The goal is not to preserve the algorithm's unconstrained choice. The goal is to define a feasible set within which the algorithm may express its information advantage.

AI-assistance disclosure: OpenAI Codex helped draft and organize the prose of this answer. The methodological choices, equations, source checks and final edits are mine.