AI Flags Hidden Risks in Crypto Futures Trading

A Bitcoin (BTC ) chart can show modest price movement while the market underneath becomes increasingly difficult to trade. Buyers may be consuming available sell orders, liquidity providers may be withdrawing quotes, and leveraged positions may be accumulating. None of those conditions is fully explained by how much the price has moved.

A new study by Suresh Rajendran and Divya Singaravelu, accepted for publication in the International Review of Economics and Finance, investigates that gap.¹ Using Bitcoin and Ether perpetual futures data, the researchers examine whether machine learning can identify short-term order-flow toxicity without directly relying on recent volatility.

The findings suggest that order-flow signals can help rank adverse-selection risk. Their broader significance is that identifying large price moves and identifying costly trading conditions are different tasks. For investors, traders, and exchange operators, understanding that difference offers a better way to judge both market quality and financial AI.

What Order-Flow Toxicity Means In Crypto Trading

Market makers supply liquidity by offering to buy and sell. They seek to earn the spread between those prices while managing inventory and the risk that prices move against them after a trade.

Consider a simplified example. A liquidity provider sells Bitcoin at an advertised price, but buying pressure pushes the market higher immediately afterward. The provider sold just before the asset became more expensive. Repeated across many trades, these adverse fills can overwhelm the income earned from spreads.

Order-flow toxicity describes trading conditions associated with that adverse-selection risk. The term does not mean that transactions are fraudulent or that the traders responsible possess confidential information. Public news, aggressive trading, liquidations, and other pressures can produce damaging flows.

Perpetual futures make this issue particularly relevant. These derivatives generally have no scheduled expiry and use funding payments to help align their prices with the underlying asset. Leverage adds another source of pressure when forced position closures send orders into the market.

How AI Analyzed Bitcoin And Ether Order Flow

The researchers studied BTC/USDT and ETH/USDT perpetual futures on Bybit using data spanning February 1, 2025, through February 16, 2026. After reserving the first 21 days for initial training, they evaluated the models across 360 out-of-sample days.

The primary model used gradient boosting, a machine learning method that combines multiple decision trees. Its 19 inputs described conditions such as spreads, order-book imbalance, trade activity, and passive order withdrawal. Direct realized-volatility inputs were excluded to test whether the model captured something beyond recent price turbulence.

Predictors used information available through the previous second. The walk-forward evaluation trained on earlier observations before scoring later ones, reducing the risk of giving the model access to future information.

The model generated a risk-ranking signal called TailScore. The primary comparison selected the highest-scoring 1% of evaluated seconds, described as a 1% gate. This was a retrospective evaluation threshold, rather than a fully specified live trading rule.

The monetary estimates above apply to the primary model at the 1% gate, using an assumed 25% fill rate. They represent modeled avoided losses after allowing for foregone spread income, not investment returns or realized trading profits.

Why Better Volatility Detection Can Produce Less Value

The most useful finding concerns the disagreement between performance measures. Simpler models based on returns, flows, or volatility generally concentrated extreme price movements more effectively across the adequately supported 1% to 5% gates. Yet the primary model without direct volatility inputs produced the highest illustrative avoided-loss point estimates for both assets.

That does not establish that it decisively outperformed every alternative. At the primary Bitcoin gate, its confidence interval overlapped those of the linear model and the model that included volatility. It does show why selecting a financial model using one impressive metric can be misleading.

A system that recognizes turbulent periods may help reduce overall exposure. A system intended to protect passive quotes must answer a narrower question: which potential fills are likely to become costly, and is avoiding them worth the income sacrificed?

This distinction extends beyond crypto. Financial AI should be evaluated against the economic decision it supports. Classification accuracy, extreme-move detection, execution costs, and profitability measure different things. Improving one does not automatically improve the others.

Why A Quiet Crypto Market Can Still Carry Risk

Recent market observations illustrate why price behavior deserves to be read alongside positioning. Coinbase Institutional’s August 2026 market positioning report described rising derivatives open interest alongside declining spot and perpetual trading volumes in July. It also reported weaker bid-side order-book liquidity.

Those observations do not validate this study’s model. They illustrate a separate point: outstanding exposure, trading activity, and available liquidity can move in different directions. A quieter market does not necessarily contain less risk.

Liquidity also has different meanings. Securities.io’s analysis of global liquidity and crypto returns concerns capital availability across the financial system. Order-book liquidity concerns the orders available to absorb transactions at particular prices. More capital in the system does not guarantee deep executable liquidity on every venue.

What This Means For Ordinary Traders

The study does not provide a ready-made retail indicator. Its practical contribution is to encourage a broader view of execution quality:

• Evaluate available market depth alongside the quoted spread.
• Distinguish rising open interest from rising trading volume.
• Assess execution costs alongside any model’s prediction accuracy.

There is also a potential tension for exchanges. If risk detection helps liquidity providers identify dangerous periods, they may reduce quote sizes or withdraw orders. That can protect their balance sheets while leaving other participants with less available liquidity. Better individual risk management does not automatically create a more resilient market overall.

What The Study Does Not Establish

The evidence covers two assets on one exchange over approximately one year. It cannot establish that the same relationships hold across other venues, smaller tokens, or different market regimes.

The study’s toxicity label combines order flow with five-second price impact. It is a constructed proxy, rather than proof of private information. After matching events on relevant market conditions, the researchers found no statistically distinguishable persistence advantage for model-selected events at the tested horizons.

The economic estimates exclude fees, slippage, funding payments, hedging costs, queue position, and overlapping or correlated positions. Live deployment would also require thresholds calibrated using only information available at the time. These omissions prevent the results from being interpreted as a demonstrated profitable strategy.

Both authors disclose that they are co-founders of Alphashots.AI, a quantitative research and trading technology firm. That affiliation is relevant context when assessing potential commercial applications.

CME Group And The Expansion Of Crypto Risk Management

The broader opportunity concerns infrastructure that lets participants trade and manage digital asset exposure. Securities.io’s coverage of real-world asset perpetuals explores how derivatives ambitions are extending beyond conventional crypto markets. As products expand, execution quality becomes increasingly consequential.

For investors interested in equity exposure to this infrastructure, CME Group offers a relevant connection through its derivatives marketplace. In June 2026, the company announced the launch of 24/7 cryptocurrency futures and options trading. It also introduced Nasdaq CME Crypto Index futures, extending its product range to diversified crypto exposure.

CME provides a way to examine the business of trading infrastructure rather than simply choosing a cryptocurrency. Relevant considerations include sustained customer participation, product adoption, competition, and the costs of supporting continuous markets. Its broader business also spans multiple asset classes.

The Bybit findings do not establish that CME uses this model or that the results transfer to its contracts. The connection is the growing importance of sophisticated risk management. Crypto market development depends on more than attracting capital: participants also need reliable ways to execute trades, hedge exposure, and evaluate what their risk tools actually accomplish.

References:

¹ Rajendran, S., & Singaravelu, D. (2026). Order-flow toxicity in cryptocurrency perpetual futures: Order flow signals, price-impact persistence, and the limits of volatility-based metrics. International Review of Economics and Finance. Advance online publication. https://doi.org/10.1016/j.iref.2026.105876