Can We Predict Market Crashes Sooner? The Hunt for New Risk-Off Signals!

A beginner-friendly summary of the verification: “Can We Predict Market Crashes Sooner? The Hunt for New Risk-Off Signals!”.

Trying to predict market downturns before they hit is a holy grail for any automated trading system (EA). Our previous research (Study 117) showed that simply looking at the magnitude of correlation between different assets wasn’t enough to reduce drawdowns. So, we shifted our focus: what if we could find a directional signal that specifically warns us of impending market drops? Our current system uses a basic filter: it reduces risk when the US500 stock index dips below its 200-day Simple Moving Average (SMA). This is like a general “market health check.” But we wanted something better, something that could act as an earlier, more precise storm warning. We explored a few candidates:

Faster SMAs: Specifically, the 100-day and 50-day SMAs, hoping they’d react quicker to price changes.
Negative Momentum: Looking for periods where prices are clearly falling.
Volatility Spikes: Sudden increases in market choppiness, similar to what the VIX index measures for stock markets.
Multiple Index Breadth: A broader measure of how many different stock indices are participating in a downturn.

How I Tested It

To see if these signals held any promise, I put them through a rigorous two-phase test:

Phase 1: Individual FX Core Performance. First, I integrated each potential signal into a single, standalone trading strategy – what we call an “FX Core.” The goal here was to see if any of these signals could improve the core’s performance, particularly its ability to avoid drawdowns (DD). Drawdown is the peak-to-trough decline in an investment, and a lower percentage is always better. I also looked at the return-to-drawdown ratio (r/DD), which tells us how much return we get for the risk we take.
Phase 2: Full System Integration. If a signal showed promise in Phase 1, the next step was to integrate it into our complete trading system (version 1.4.1), which usually involves multiple FX Cores or strategies running simultaneously. This is the ultimate test, as individual improvements don’t always translate to the overall system.

What Happened?

Phase 1: Promising Results for Individual Strategies!

The good news arrived almost immediately when testing the faster SMAs. These signals actually worked quite well for the individual FX Core!

Steady Improvement with Faster SMAs: When we switched from the 200-day SMA to the 100-day, and then to the 50-day SMA as our risk-off trigger, we saw a consistent, “monotonous improvement” in our core’s drawdown and risk-adjusted return.
Drawdown (DD): Improved from -7.4% (with SMA200) to -6.9% (SMA100), and finally to -6.5% (SMA50). In other words, the individual strategy experienced smaller peak-to-trough losses!
Return-to-Drawdown (r/DD): Increased from 0.24 to 0.25 to 0.27. This means we were getting a slightly better return for the level of risk we were taking. This improvement wasn’t just a fluke. The faster SMAs seemed to catch market downturns earlier, and this wasn’t simply a case of “overfitting” to past data; it suggested a genuine directional signal was at play.
Other Candidates Fell Short: Unfortunately, the other signals we tested – negative momentum, volatility spikes, and multi-index breadth – didn’t meet our performance standards for the individual core.
Mixed Bag in Different Market Conditions: Interestingly, these faster signals showed their strength during actual or simulated crisis periods (like 2015-2018 and our future-simulated 2024-2026 data). However, they sometimes “cut into” profits during strong, clear uptrends (like 2021-2024) because they got us out of positions too early. So, it was a mixed bag, improving in 2 out of 4 tested periods.

Phase 2: System-Wide Disappointment

Here’s where the excitement hit a wall. Despite the clear improvements for the individual FX Core, those gains simply did not translate to the full trading system (v1.4.1).

No DD Improvement: When we implemented the faster SMA50 into the complete system, the overall system drawdown was 9.7%, which was almost identical to our baseline of 9.6% (which uses SMA200). Even worse, using SMA100 actually increased the system’s drawdown to a painful 10.3%!
The Familiar Problem: This is a recurring theme we’ve encountered before (in Studies 51 and 86): “improvements in an individual core don’t necessarily lead to improvements in the entire system.”
Why the Disconnect? The core reason is that the system’s overall drawdown is primarily determined by correlation drawdown between sleeves. Think of “sleeves” as the different individual strategies or asset classes within your larger system. Even if one individual strategy (an FX Core) ducks for cover faster, the overall system’s drawdown wall remains firmly in place if the other strategies are still correlated and falling together. It’s like patching one leak in a boat that has multiple, interconnected holes – the boat still takes on water.
No Re-Leveraging: This means we can’t “re-leverage” or increase our position sizes based on this perceived improvement, as the system’s overall risk hasn’t genuinely decreased.
Monte Carlo (MC) Slightly Worse: Even our Monte Carlo simulations, which test system robustness under various scenarios, showed a slight decrease, indicating the faster SMAs didn’t make the system more resilient.

What I Learned

While this research didn’t deliver the system-wide drawdown reduction we hoped for, it wasn’t a total loss.

A “Safety Valve” for Aggressive Strategies: The only real win was for monthly worst-case drawdown (M1). The SMA50 trigger reduced M1 from 2.42% to 1.69%. This means that while it didn’t help the overall system’s biggest dips, it did a better job of reining in those painful intraday givebacks or “spills” when things got rough. This makes the SMA50 useful as a “safety valve” for our more aggressive FX Cores or when we’re running higher leverage. It’s a similar concept to a “scale-out” strategy, where you exit a position in parts to reduce risk. However, for the standard v1.4.1 system, we’ll be sticking with the SMA200.
The Unbreakable Wall of Correlation: This study, along with previous ones (116-118), really hammered home a crucial point: the system’s overall drawdown – the true constraint on our profits – is primarily determined by correlated drawdowns between its different sleeves. You simply cannot reduce system-wide drawdown by making single-sleeve risk-off adjustments, whether it’s through faster SMAs, looking at the magnitude of correlation, or even using scale-out techniques.
The Path Forward: To truly reduce system drawdown, we need “truly uncorrelated positive expected value (EV) sleeves.” In other words, we need strategies that are expected to be profitable over the long run, but crucially, their ups and downs are completely independent of each other. And the hard truth is, these truly uncorrelated opportunities are likely not found within the price data of existing, commonly traded assets. The only real path to genuinely lower system drawdown seems to be through exploring entirely new data sources or completely different asset classes. This research reconfirms that our current v1.4.1 system is already operating at the “drawdown frontier” – meaning, with its current components, it’s as good as it gets in terms of balancing risk and reward. It’s a humbling but important realization for any EA developer!

How this connects

This verification builds on earlier ones (what failed before and what I tried this time, comparisons between approaches).