If 2008 Came Again — We Stress-Tested Our Defense Against Real Data
Our backtest contains only one major crisis: COVID. But COVID was a V-shape — a sharp drop followed by a sharp rebound. What if a 2008-style crisis came, one that grinds down for a full year? We pulled 2005 data via CYBOS and actually ran it.
Our regime-adaptive strategy (REGIME_C) has a strong 12-year backtest. But within those 12 years (2015–2026), there was only one major crisis: COVID (2020). And COVID was a V-shape — a one-month plunge followed immediately by a sharp recovery.
That is where the problem begins. The 2008 global financial crisis was a completely different animal. It ground downward for a full year, from peak to trough. Our strategy has never once faced a crisis like that. With the KOSPI sitting in record territory around 8,000 (close of 8,051 on 2026-07-07), leaving this untested regime unexamined looked like our single largest risk.
This post is not a forecast. It is not about where the market is going. It is a record of what the data says about how our system would have behaved had it met a 2008-style crisis.
Step 1: Ask the statistics first — "how bad was the worst?"
Before obtaining real 2008 data, there is a question we can answer with what we already have. If we randomly reshuffle the order of our strategy's monthly returns, how deep can the drawdown get? (Block bootstrap, 20,000 runs.)
| Scenario | Max drawdown |
|---|---|
| Actually realized | −21.8% |
| Simulation median | −28.0% |
| 5th percentile (bad case) | −43.8% |
| 1st percentile (very bad case) | −51.8% |
The realized −21.8% sat at the favorable edge of the possible-drawdown distribution. Same strategy, same returns — but had the order been unlucky, −44% to −52% was entirely possible. The low realized drawdown mixed skill with path luck.
Step 2: Actually run 2008
Statistical estimates were not enough. We needed to see whether our defensive switch actually fired during the real 2008 window. For that, we obtained 2005–2013 prices via CYBOS.
We started with a reliability check. We replicated our regime-classification logic and ran it over 2015–2026; the output matched our live records for all 137 months (100%). Having confirmed the logic was faithfully replicated, we applied the same logic to 2007–2009.
Good news and bad news arrived together.
- ✅ The switch fired on time. Just one month after the KOSPI peak (2007-10-31), the regime flipped to BEAR. Detection itself was fast.
- ❌ But the defense failed. While the KOSPI fell −54.5% through 2008, our strategy's defense curve applied to that window was −47% to −54%. It barely cushioned anything.
Why did the switch fire but the defense fail?
The cause was a single rule. Our strategy has a rule: "when price falls well below its 60-day average = it has fallen too far, so a rebound is near (oversold)" — and it releases the defense to ride market-representative stocks.
- In COVID (2020) this rule was right. After the plunge, a real rebound came.
- In 2008 it was exactly backwards. Price stayed below the average for a full year, so the "oversold" signal was on for the entire decline. In 8 of the crisis's 10 months the system was in this state — betting on a rebound all the way down, and falling with the market.
This rule was tuned to a single V-shaped sample (COVID, ~15 months). It did not generalize to a slow, grinding 2008-style decline.
Combining the two answers
| Question | Statistical estimate | 2008 empirical |
|---|---|---|
| Worst drawdown? | 1st pct −52% | proxy −47% to −54% |
| Does the defense generalize? | untested | refuted |
| What was the realized −21.8%? | open | path luck of COVID's V-shape |
The statistics' warning of a "−52% tail" and the 2008 empirical confirmation of "defense failure" pointed the same way. Our strategy's −21.8% drawdown was not a floor set by skill — it was a favorable ceiling set by luck.
What this means for us
This verification changed our own perception of risk.
We now size our strategy's tail risk not at the realized −22%, but at the −50%-range that both the bootstrap and the 2008 empirical jointly point to. This is not a prediction that the market will crash. It is a commitment to perceive our own system's properties conservatively, and to impose operating discipline on ourselves accordingly — no leverage, and a cash buffer.
The most dangerous sentence in quant is "this strategy is robust in a crisis." Our basis for believing that was a single event (COVID), and that one event happened to be the easiest possible form to defend. Pushing the data far enough to break our own illusion — that is why we publish this.
Limitations (stated plainly)
The 2005–2013 window lacks per-stock supply-flow data and point-in-time index membership. So "exactly which stocks were bought and how much they earned back then" cannot be fully reproduced, and the defense curve above is a proxy built from regime classification and approximate exposure. That said, the defense-failure conclusion holds identically under both optimistic and pessimistic assumptions, so the proxy's uncertainty does not overturn it.
Figures here are based on 2015–2026 backtests (0.35% round-trip cost, survivorship-adjusted) and a 2007–2009 regime-classification reproduction (price-based proxy). Everything is an observation about our system's past and simulated behavior — not a forecast of market direction, nor a recommendation to buy or sell any security or asset. Backtest and simulation results do not guarantee future returns.