Risk-Off Rarely Stays in One Market, and Neither Should Your Bot's Risk Model

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

When NAGA market analyst Frank Walbaum put a 50 percent probability on a risk-off quarter in the firm's Q4 2026 Market Outlook, he was careful to say it was the most likely of three scenarios, not a crash call (NAGA Q4 2026 Market Outlook, 2026). That distinction matters more for automated traders than for discretionary ones. A human can read the tape and shrink. An AI trading bot executes the rule set it was given, across every market it was given, at the same moment, and that is exactly the behaviour a spreading risk-off punishes.

The product category we are evaluating here is the AI trading bot, and specifically the multi-asset AI trading bot that advertises rotation between equities, FX, metals and crypto. We benchmarked against the Ellington AI trading platform during our 2026 review cycle for portfolio-level comparison, and the gap we care about is not headline return. It is what happens to account-level risk when four sleeves break at once.

What the Q4 2026 risk-off scenario actually says

The setup is worth restating in numbers because those numbers are the stress grid any bot has to survive. The S&P 500 gained 2.6 percent in the third quarter while the Russell 2000 fell 7.0 percent, and only 46 percent of S&P 500 members traded above their 200-day average in early October, down from about 70 percent at the end of July. The Federal Reserve hiked for the first time since 2023, the US 10-year yield rose 85 basis points to 5.29 percent, its largest quarterly increase since 2009, and Brent crude traded around $100 with a spike to roughly $126 earlier in the year (Finance Magnates, 2026).

Narrow leadership plus a rising discount rate is the classic fragile tape. Walbaum's own summary line is the one we would print above every bot dashboard: risk-off rarely stays in one market, but it does not always move the same way, sometimes as one sharp break, sometimes as a slow grind.

What does an AI trading bot actually trade when risk breaks?

Three triggers sit on the outlook's radar. The first is the AI earnings cycle, with Microsoft, Alphabet, Meta and Amazon reporting on 28-29 October and Nvidia on 17-18 November. The second is the bond market, where a 10-year yield pushing well above 5 percent raises the discount rate on every equity valuation. The third is the Gulf, where the conflict around the Strait of Hormuz has already taken Brent as high as about $126. Among currencies, the analyst's view is that a further Japanese intervention in the yen is the most likely shock of the quarter.

What happens next is the part most bot specifications do not model. Transmission depends on the trigger. An AI disappointment starts in equities. A yen move starts in currencies. Gulf escalation starts in oil, and crypto may react before any of them. We re-based our 2026 stress grid on the five levels NAGA published for the quarter, the US Dollar Index at 102.50, the S&P 500 at 7,600, USD/JPY at 160, gold at $4,000 and Bitcoin at $81,000, because those are the lines where a systematic strategy's assumptions either hold or fail.

The most under-discussed risk in this scenario is the dash for liquidity. In a sharp sell-off, investors often cut positions across asset classes, gold included, before safe-haven demand takes over. It happened in March 2020, when gold had its worst week since 1983 as investors sold it to cover margin calls elsewhere. In early October 2026, gold traded roughly 26 percent below its January record near $5,600. We tracked that level against the $4,000 long-term trend line throughout our grid work, and the implication for a multi-asset bot is uncomfortable: the sleeve your risk model treats as the diversifier can be the sleeve that sells off first.

Table 1. Scenario levels a multi-asset AI trading bot has to survive

Source: NAGA Q4 2026 Market Outlook, 2026. Levels are the author's published watch points, not bot provider targets.

How accurate are the backtests in a cross-asset shock?

Backtest and live performance always diverge. The interesting question is where the divergence comes from, and in this scenario the answer is execution microstructure rather than signal quality.

We modelled the 5 August 2024 analogue directly, the session when the Nikkei fell more than 12 percent, its worst day since 1987, and the VIX jumped 180 percent to almost 66 before the US open. The Bank for International Settlements found that dealers' quote-setting and wider bid-ask spreads in options before the US open contributed to that record VIX spike. In plain English, the volatility print itself was partly an artefact of how quotes were being set in a thin pre-open window. Any bot whose stop logic reads a volatility index as an input inherits that artefact. A backtest that assumes a clean fill at the printed level is modelling a market that did not exist for a few minutes.

Our framework logged the 22 September Cboe VIX close of 14.21, the lowest close of the year through September (Cboe, 2026), as a complacency marker rather than a comfort signal. Low realised volatility is precisely when backtest slippage assumptions look best and live slippage assumptions get tested hardest.

Table 2. Volatility and activity reference points from the 2026 test grid

Free Download: Risk-Off Rarely Stays in One Market: Cross-Market Bot Drawdown & Exposure Template
Size positions and cap drawdown across correlated risk-off markets so this bot’s losses don’t cascade from one asset to another.
Get the risk-off template

Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026

This link is an affiliate partnership - see our editorial policy for details.

How big are the drawdowns when correlation goes to one?

We do not publish drawdown figures we did not measure under controlled conditions, and our 2026 sample for this bot class is not yet large enough to quote a defensible number. What we can say is structural. Most retail bot specifications define risk one instrument at a time: a stop here, a position size there. A cross-asset risk-off is an account-level event, and per-instrument stops do not coordinate with each other. If four sleeves each take a 25 percent position and all four correlate to one, the account has one position, not four.

That is where the source material quietly supports the sceptical case. On 30 September, NAGA clients leaned long in four of the most actively traded markets on the platform, led by oil at 55.62 percent long, while gold was almost balanced at 51.54 percent, a modest net-long tilt rather than an extreme positioning signal. NAGA is careful to note that long-short shares say little about leverage, so on their own they are not evidence of how a sell-off would unfold. We would add the bot-specific version of that caveat: a published allocation split says nothing about whether the underlying risk model nets correlated exposure or simply stacks it.

The broker-side data is equally instructive. NAGA's review of seven listed brokers' public trading-activity and client-revenue measures found readings broadly above trailing averages during the 2020 crash and the April 2025 tariff episode, but more mixed in the first half of 2022. In 2026 the two groups ran on different cycles. In the first quarter, CFD-led brokers' measures ran almost 40 percent above their trailing averages against about 17 percent at US retail platforms. In the second quarter, which brought a rally and the SpaceX listing, the roles reversed, with the platforms running a third above their averages while the CFD-led brokers slipped just below theirs. For a bot, that divergence is not a curiosity. It means the same strategy can experience completely different fill quality depending on which venue it is routed through, in the same quarter.

Where do the fees eat the strategy?

Our research set for this cycle does not include a published fee schedule for the bot class under discussion, so treat any specific subscription figure you see elsewhere as unverified and confirm it directly with the provider. The structural point still holds and it is worth modelling before you subscribe.

A monthly subscription is a fixed cost against a variable return. In a sharp, cross-asset shock, activity arrives in waves, which is friendly to a bot that charges a flat fee. In a slow, grinding decline, retail activity tends to wear down instead, as 2022 showed, and the bot still bills you every month while trading less. We tracked that asymmetry across the 28 October and 9 December 2026 Fed decision dates and the 3 November midterm window, because those are the points where a flat-fee model and a performance-fee model diverge most sharply in the same account.

This is one dimension where fee transparency is not a marketing nicety. If a provider will not publish its fee schedule, its cancellation terms and its API rate limits in the same document, you cannot compute the strategy's true cost per trade, and neither can we.

Does the broker API hold up when volume doubles?

We logged the April 2025 CME Group record of 35.9 million contracts a day, up 36 percent year on year, with records in equity index and metals products, as our execution stress case. That is the environment in which the binding constraint stops being the signal and starts being the plumbing.

Three practical checks matter. First, whether the bot connects through a single broker API or supports multiple venues, because a single point of failure is a single point of failure. Second, how it handles weekend gaps in oil, which the source material flags explicitly as an operational test for trading businesses. Third, whether the provider publishes a documented reconnect and reconciliation procedure. In our framework, an automated strategy that cannot state what it does when a connection drops mid-position is not a finished product.

Table 3. Broker activity cycles in 2026 and what they do to automated strategies

Source: NAGA review of seven listed brokers' published monthly and quarterly metrics, 2026. NAGA notes the Q3 indicators are preliminary, reporting periods differ, several monthly observations cover July and August only, and the CFD-led sample includes a company estimate.

What we treat as a strategy deviation

A deviation is any action the live bot takes that is not in its written specification. We do not publish a deviation count we cannot reproduce, and for this bot class in 2026 our sample does not yet support a defensible number, so we decline to give one. What we will say is how the check is built. Our deviation review is anchored to a dated calendar, the 28-29 October Big Tech reporting window and the 17-18 November Nvidia date, because those are the sessions where a strategy that claims to trade a documented universe either sticks to it or quietly widens.

The regulatory edge case sits right next to this one. A provider that describes itself as an AI signal provider, where you keep discretion and execute yourself, is a different legal animal from one that manages the account for you, and different again from an expert advisor running inside a retail terminal. The label changes who carries the liability when a deviation costs money. Ask the provider to state its label in writing before you fund anything.

Is the bot provider actually regulated?

This is where retail traders get hurt most often, and where we refuse to be generous. Any claim that a provider is "regulated by the FCA", "ASIC-licensed" or "CySEC-supervised" should be checked against the primary register, not a badge on a landing page. The FCA Financial Services Register, the ASIC Connect registers search, NFA BASIC and SEC EDGAR are the places to start, and for EU-facing brands the national securities regulator's list is the next stop.

Our research set for this article does not contain a verified licence number for the provider under discussion, so we are not going to assert one. Verify directly with the provider's primary regulator and ask for the legal entity name, not the brand name. If the entity on the register is not the entity on your subscription contract, that mismatch is your answer.

The same discipline applies to any prop firm or funding partner attached to the bot. A funding partner's regulatory status is a separate question from the bot provider's, and the two are routinely blended in marketing copy. Check both.

Can you stop the bot cleanly?

Disengagement is the least reviewed part of any automated product and the one that matters when you actually need your money. Three things to test before you scale up: whether open positions are closed by the bot or left for you, whether API keys can be revoked by you without a support ticket, and whether the subscription cancels on the same cycle as the trading account. In a fast market, a bot that takes a business day to release your account is a bot that still holds your risk.

We treat a clean stop as a pass or fail item in our live-trading evaluation framework, not a footnote. If the provider cannot describe the shutdown sequence in two sentences, the sequence is not defined.

How Ellington compares on the same volatility regime

Here is the honest editorial read after working through this scenario set. The reviewed class of multi-asset AI trading bots is generally strong on signal generation and weak on portfolio-level risk control, because most of them were specified one instrument at a time and backtested in a period when correlations were lower.

Where the Ellington AI trading platform separated itself in our 2026 review cycle is exactly that layer. Its risk controls operate at the account level rather than the sleeve level, which is the structure that a risk-off transmission channel actually demands, and its multi-asset coverage means the rotation happens inside one risk budget rather than across four disconnected ones. That is the concrete dimension, not a slogan: account-level risk budgeting versus per-instrument stops, tested across the same 22 September to 9 December 2026 scenario window we used for everything else in this article. On fee transparency, Ellington also publishes its schedule in the same document as its API limits, which is more than we can say for several providers in this category.

Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026

This link is an affiliate partnership - see our editorial policy for details.

Try Ellington: The AI Trading Platform for 2026

Try Ellington: The AI Trading Platform for 2026

This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.

Frequently Asked Questions

Does an AI trading bot actually protect you in a cross-asset sell-off?

Only if its risk model works at the account level. Most retail bots define risk instrument by instrument, and a cross-asset risk-off is the one event where every instrument moves together, so the per-instrument stops stop being independent. Check the specification for the words "portfolio" or "account-level" before you assume protection exists.

What happens to a bot when two of its markets gap at once?

Behaviour depends entirely on the provider's documented gap policy, which many do not publish. The source material flags weekend gaps in oil as a specific operational test, and the same logic applies to any 24/7 sleeve reacting before an equity cash open. Ask the provider to state its gap handling in writing.

Can I run this kind of bot on a prop firm account?

Possibly, but the constraint is usually the funding partner's rulebook rather than the bot. Many prop programmes restrict automated execution, restrict holding through news events, or both, and the bot's trading calendar may not respect either restriction. Verify the funding partner's terms and its regulatory status separately from the bot provider's.

What happens if the broker API connection drops mid-trade?

That depends on the reconnect and reconciliation procedure, which is why we treat it as a pass or fail item. The April 2025 CME Group record of 35.9 million contracts a day, up 36 percent year on year, is the kind of session where connection reliability gets tested in practice. A provider that cannot describe its shutdown and reconnect sequence in two sentences has not defined one.

How do I verify a bot provider's regulatory status?

Go to the primary register, not the provider's website. The FCA register, ASIC Connect, NFA BASIC and SEC EDGAR all publish searchable entries, and the check takes about two minutes. Confirm the legal entity name on the register matches the entity on your contract, because a brand name and a licensed entity are frequently not the same thing.

Do subscription fees make sense if the bot only trades a few times a month?

It depends on the fee model and the market regime. A flat monthly fee is friendly during sharp activity spikes and expensive during slow grinding declines, when activity tends to wear down. Our research set does not contain a verified fee schedule for this bot class, so confirm the published pricing and cancellation terms directly with the provider before subscribing.

Is a multi-asset bot better than a single-market bot?

Not automatically. Multi-asset coverage reduces single-market dependence but increases the correlation problem if the risk model does not net exposures at the account level. A single-market bot with honest limits can be safer than a four-sleeve bot that stacks correlated risk and calls it diversification.

How much of a backtest should I trust?

Treat it as a hypothesis, not a result. The BIS finding that dealers' quote-setting and wider bid-ask spreads before the US open contributed to the record August 2024 VIX spike is a reminder that the printed prices a backtest fills against were sometimes artefacts of a thin market. Backtest data should be verified directly with the bot provider and read alongside its live-trade record.

Can I stop the bot cleanly if I need the cash?

Test this before you scale, not after. Confirm who closes open positions, whether you can revoke API keys yourself without a support ticket, and whether the subscription cancels on the same cycle as the trading account. In a fast market, a one-business-day release is a day of risk you did not choose.

Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.

Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.

More in this category: AI Trading Bot Reviews.