top of page
Search

News VIX Part 2: Stress-Testing the Model in a Crash Regime

  • Writer: 오리 오리
    오리 오리
  • 16 hours ago
  • 8 min read

1) What changed since Part 1

In Part 1 (late Jan → mid-Feb 2026), I introduced News VIX—a headline-based volatility gauge—and showed it could function as a usable "big-move radar" for next-day KOSPI moves. The results were encouraging: 63.6% accuracy on volatility classification, and an 80% hit rate on directional tilts (with caveats about sample size).

This post is Part 2: a larger window (late Feb → mid-Mar 2026), with 18 observations instead of 11. But the market regime was far more extreme—and the results are more humbling. The honest summary: the model's volatility accuracy dropped to 50%, and the directional layer essentially broke.

That's not a failure to hide. It's the most informative kind of test—one that reveals exactly where and why the framework's assumptions stop holding.


2) Market backdrop: what was happening in KOSPI

This window (late Feb → mid-Mar 2026) wasn't just volatile—it was structurally different from Part 1. The moves were larger, more clustered, and less predictable from headlines alone.

Key events:

  • Feb 24–26 ("6000 breakout rally"): KOSPI surged from ~5846 to 6307 over three sessions (+7.9% cumulative), driven by semiconductor optimism and foreign inflows. Headlines were overwhelmingly positive.

  • Feb 27–28 (reversal): KOSPI dropped from 6244 to 5791, a single-day decline of –7.24%. Headline tone the day before was not particularly alarmed—News VIX sat at just 1.67.

  • Mar 4 ("Black Tuesday"): The biggest single-day move in this dataset—KOSPI fell –12.06%, from 5791 to 5093. News VIX the day before was only 1.27.

  • Mar 5 (snap-back): A +9.63% rebound the next session, with News VIX at 1.80 (alert zone, barely).

  • Mar 6–7 (whipsaw continues): After a flat session (despite VIX 3.33—the highest reading in the entire sample), KOSPI dropped –5.96% on Mar 7.

  • Mar 10 (another rebound): +5.35%, again from a non-alert News VIX reading (1.67).

  • Mar 17–18 (late rally): KOSPI surged +5.04% from 5640 to 5925—completely unannounced by headline intensity (VIX was just 1.27).

  • Mar 18–19 (pullback): –2.73%, this time correctly flagged by News VIX at 2.60.

The KOSPI range during this window: 5,093 → 6,307—a span of over 1,200 points. Nine out of 18 trading transitions exceeded 2% in absolute terms. This is not a normal market. It's a crash-and-recovery regime where the question is not "will there be a big day?" but "which days won't be big?"

3) News VIX distribution and threshold check

The distribution of News VIX readings in this window is wider than Part 1:

  • Range: 0.47 to 3.33 (vs. 1.00 to 2.67 in Part 1)

  • Mean: 1.64

  • Median cluster: still around 1.3–1.8

The same operational thresholds from Part 1 were applied:

  • Calm: below ~1.7 → "tomorrow likely smaller move"

  • Alert: ~1.7+ → "big move becomes more plausible"

  • High alert: ~2.0+ → "risk expansion mode"

The problem: in a regime where 50% of days were "big," the base rate of large moves was so high that the calm zone became unreliable. Five big days occurred when News VIX was below 1.7—including the two largest moves in the entire dataset (–12.06% and +9.63%-adjacent).

This suggests the thresholds need to be regime-adaptive: in a sustained high-volatility environment, "calm" should not mean "safe."

4) Accuracy: volatility classification

Same methodology as Part 1: a "big day" is a next-day move exceeding 2% in absolute terms. News VIX ≥ 1.7 = alert.

Results (18 observations):

Metric

Part 1 (11 obs)

Part 2 (18 obs)

Overall accuracy

63.6%

50.0%

Precision

60.0%

50.0%

Recall

60.0%

44.4%

True Positives

3

4

False Positives

2

4

False Negatives

2

5

True Negatives

4

5


The drop is real and meaningful. The core issue is the 5 false negatives—big days that News VIX failed to flag. These include:

  • 2/27 → 2/28 (–7.24%): VIX was 1.67, just below threshold

  • 3/3 → 3/4 (–12.06%): VIX was 1.27—completely missed

  • 2/23 → 2/24 (+2.11%): VIX was 1.00

  • 3/9 → 3/10 (+5.35%): VIX was 1.67, borderline

  • 3/17 → 3/18 (+5.04%): VIX was 1.27

Two patterns emerge from the misses:

  1. Crash aftershocks are invisible to headlines. The –12.06% day and the +5.35% rebound both came from "calm" readings. In an ongoing crisis, the market can produce extreme moves without new headline catalysts—positioning, margin calls, and flow dynamics take over.

  2. Borderline readings (1.6–1.7) keep appearing before big days. Two of the five false negatives had VIX at exactly 1.67. The 1.7 threshold may be too tight for this regime

5) Accuracy: directional layer

In Part 1, I reported an 80% hit rate on directional tilts (4/5). That number was always marked as preliminary, and this window confirms why.

Method: I classify headline tone as "upside tilt" when the negative article share is below 20%, "downside tilt" when above 30%, and "neutral" otherwise.

Results: Of 18 observations, 11 had a non-neutral tilt. Of those 11, only 3 matched the next day's direction (27.3%).

That's worse than a coin flip. The directional layer didn't just underperform—it actively misled.

Why it broke:

  • Positive tone ≠ next-day up in a crash regime. On 3/6, headlines were overwhelmingly positive (negative share 12.5%), but KOSPI fell –5.96% the next day. Post-crash headline coverage tends to be optimistic (recovery narratives, policy support), but the market is still deleveraging.

  • Negative tone ≠ next-day down during recoveries. On 3/9 (negative share 28.9%) and 3/17 (32.3%), the market surged +5.35% and +5.04% respectively. Headlines lagged the bottoming process—they were still writing about damage while the market was already bouncing.

  • Tone follows the market; it doesn't lead it. This is the key lesson. In a trending or slow-moving market, headline tone can slightly anticipate direction. In a crash-and-recovery regime, tone is a lagging indicator—it describes yesterday's move, not tomorrow's.

The honest conclusion: the directional layer needs a fundamentally different design for crisis regimes—possibly incorporating rate-of-change in tone rather than absolute levels, or using a "tone reversal" signal instead.


6) Day-by-day post-mortem (prediction vs. what happened)

2/23 → 2/24 (News VIX 1.00; negative share ~34%) Signal: Calm; tone leaned negative. Outcome: +2.11% (big up day) Read: Double miss—intensity was low and tone pointed down, but the market rallied. This was the start of the "6000 breakout" move, driven by semiconductor flows that headlines hadn't yet priced into their intensity. A reminder that when a rally ignites from positioning rather than narrative, headline-based tools will be late.

2/24 → 2/25 (News VIX 2.47; negative share ~17%) Signal: High alert + strongly constructive tone. Outcome: +1.91% (moderate up, just below 2% threshold) Read: Volatility call was a false positive—the move wasn't technically "big" by the 2% cutoff, but it was close. The upside tilt was directionally correct. In practice, a trader hearing "high alert + upside lean" would have been reasonably well-positioned. This is a case where the binary threshold penalizes what was actually a useful signal.

2/25 → 2/26 (News VIX 2.53; negative share ~7%) Signal: High alert + overwhelmingly positive tone. Outcome: +3.67% (big up day) Read: Clean hit on both layers. Intensity correctly flagged a large move, and the extreme positive skew in headlines matched the upward direction. This was the peak of the rally—KOSPI hit 6307. The model works well when headline consensus and market momentum are aligned.

2/26 → 2/27 (News VIX 1.53; negative share ~25%) Signal: Calm; neutral tone. Outcome: –1.00% (small down) Read: Correct. The calm regime matched a modest pullback. No signal, no drama.

2/27 → 2/28 (News VIX 1.67; negative share ~20%) Signal: Calm/borderline; neutral tone. Outcome: –7.24% (massive down day) Read: Major miss. This is the single most important failure in the dataset. News VIX at 1.67 is just 0.03 below the alert threshold, and the next day produced a –7.24% crash. Headlines the day before were not alarmed—the reversal came from positioning unwind and external catalysts that headlines hadn't yet absorbed. This day alone makes the case for lowering the alert threshold in high-vol regimes.

3/3 → 3/4 (News VIX 1.27; negative share ~29%) Signal: Calm; mildly negative tone. Outcome: –12.06% (largest single-day drop in the dataset) Read: The worst miss. A –12% crash from a "calm" reading. The negative tone was elevated but not extreme, and headline intensity was unremarkable. This is the clearest example of a structural limitation: in a cascading sell-off driven by margin calls, forced liquidation, and global contagion, new headlines simply don't appear fast enough to flag the risk in advance. The crash was mechanical, not narrative.

3/4 → 3/5 (News VIX 1.80; negative share ~31%) Signal: Alert (barely); downside tilt from elevated negative share. Outcome: +9.63% (massive rebound) Read: Intensity was correct—a big move did follow. But direction was wrong: the model leaned negative while the market staged a historic snap-back. This is classic post-crash behavior: headlines are still processing damage while the market is already pricing recovery. The volatility layer earned its keep here; the direction layer did not.

3/5 → 3/6 (News VIX 3.33; negative share ~18%) Signal: Extreme alert (highest reading in entire dataset) + positive tone. Outcome: +0.02% (effectively flat) Read: The most dramatic false positive. News VIX screamed "huge move incoming," but the market did almost nothing. Why? After a +9.63% rebound, both buyers and sellers were likely exhausted. Extreme headline activity doesn't always translate to market movement—sometimes it reflects the media processing what just happened rather than signaling what comes next. This is the "headline echo" problem: after a shock, coverage volume spikes but forward-looking information content drops.

3/6 → 3/7 (News VIX 2.47; negative share ~13%) Signal: High alert + upside tilt (low negative share). Outcome: –5.96% (big down day) Read: Intensity was correct (big move followed), but direction was completely wrong. Headlines were optimistic—recovery and policy support narratives dominated—but the market resumed selling. In a whipsaw regime, yesterday's optimism is tomorrow's trap. The positive tone was describing the rebound that already happened, not predicting the next leg down.

3/9 → 3/10 (News VIX 1.67; negative share ~29%) Signal: Calm/borderline; mildly negative tone. Outcome: +5.35% (big up day) Read: Another borderline miss on volatility (VIX at exactly 1.67 again), and direction was wrong. The negative tone reflected ongoing crisis coverage, but the market had found a floor and was bouncing. This is the third time a reading of 1.67 appeared before a big day—strong evidence that 1.7 is too high as a threshold.

3/10 → 3/11 (News VIX 0.60; negative share ~38%) Signal: Deep calm; negative tone (but low article count—only 20+12=32 total). Outcome: +1.40% (small up) Read: Correct on volatility—no big move. The elevated negative share came from a very small sample of articles, which makes the ratio unreliable. On days with low total article volume, negative share should be down-weighted.

3/11 → 3/12 (News VIX 1.40; negative share ~17%) Signal: Calm; mildly positive tone. Outcome: –0.48% (small down) Read: Correct. Calm regime, small move. The mild tone mismatch (positive lean vs. slight down) is within noise.

3/12 → 3/13 (News VIX 1.93; negative share ~25%) Signal: Alert; neutral tone. Outcome: –1.72% (moderate, not big) Read: False positive on volatility—the alert zone triggered but the move was under 2%. However, –1.72% is not trivially small either. As with 2/24→2/25, the binary cutoff may be too rigid here. In practice, "alert" + a –1.72% move is a reasonable outcome for risk management purposes.

3/13 → 3/14 (News VIX 2.00; negative share ~23%) Signal: Strong alert; neutral tone. Outcome: +1.14% (small up) Read: False positive. The market had stabilized around 5500 after the crash, and headline intensity was reflecting the prior week's turbulence rather than new forward-looking risk. This is another instance of the "echo" problem—elevated News VIX readings in the aftermath of a crisis overstate the next-day risk because they're measuring backward-looking coverage.

3/16 → 3/17 (News VIX 0.87; negative share ~31%) Signal: Calm; downside tilt. Outcome: +1.63% (small up) Read: Correct on volatility (no big move), wrong on direction. The negative tone reflected lingering pessimism, but the market had quietly shifted into a recovery phase. A small miss on direction, but the volatility call was right—which is what matters most.

3/17 → 3/18 (News VIX 1.27; negative share ~32%) Signal: Calm; downside tilt. Outcome: +5.04% (big up day) Read: Major miss. Another large rally that News VIX completely failed to anticipate. The negative tone was high, and yet the market surged. This day, like 3/3→3/4, shows that some of the biggest moves come when headlines are telling a different story. The rally was likely driven by technical factors (short covering, oversold bounce) that are invisible to headline analysis.

3/18 → 3/19 (News VIX 2.60; negative share ~11%) Signal: High alert + strongly positive tone. Outcome: –2.73% (big down day) Read: Intensity was correct—a big move followed. Direction was wrong (positive lean vs. down move). After a +5% rally, the market pulled back, but headlines were still celebrating the surge. Same pattern: tone lags, especially at turning points.

3/19 → 3/20 (News VIX 1.40; negative share ~25%) Signal: Calm; neutral tone. Outcome: +0.31% (flat) Read: Correct. Calm regime, negligible move. A quiet end to a turbulent month.


 
 
 

Comments

bottom of page