Forecasting Next-Day KOSPI Moves with a Headline-Based News VIX

오리 오리
18 hours ago
6 min read

1) Why I built a “News VIX”

KOSPI prices are numbers, but short-horizon moves (1–2 trading days) are often driven less by fundamentals and more by compressed psychology: uncertainty, fear, and conviction. Headlines can change risk perception faster than earnings models can.

That’s why I built News VIX—a headline-based volatility gauge designed to answer one practical question first:

“Will tomorrow be a big day?”(a large move, regardless of whether it’s up or down)

This framing is intentional. VIX-type indicators are typically more reliable at detecting risk expansion (move size) than predicting direction. So I treat News VIX as a next-day big-move radar, and then layer directional hints on top using headline tone.

2) What News VIX represents (no formulas, just intuition)

My daily workflow compresses headlines into two types of signals:

A) News VIX = “Move size / uncertainty intensity”

When News VIX rises, it means headlines are filled with shock, policy uncertainty, risk-off language, and macro sensitivity. In those conditions, the market is more likely to deliver outsized moves the next day—either a sharp drop or a sharp rebound.

B) Headline tone = “Directional tilt (probabilistic)”

Alongside intensity, I track how negative the day’s headlines are and whether positive vs. negative coverage is clearly imbalanced.

When the negative share spikes, it often reflects risk-off pressure and a higher chance of downward tilt.
When positive tone dominates strongly, it often adds an upward tilt.

Important: this is not a guarantee. It’s a tilt, not a daily up/down machine—especially during macro shock regimes.

3) Market backdrop: what was happening in KOSPI during this window

This window (late Jan → mid-Feb 2026) is a high-volatility regime where sentiment and positioning flipped quickly:

Feb 2 (“Black Monday”): KOSPI fell about 5.26%, breaking the 5000 level amid a major shock narrative (often described as “Wash shock”).
Feb 3: The market snapped back with a +6.84% surge, setting a new closing high narrative.
Feb 5: Another risk-off wave hit—KOSPI dropped -3.86%, driven by US tech weakness and record-scale foreign selling headlines.
Feb 12 (“5500 era”): KOSPI closed above 5500 for the first time, supported by semiconductor momentum and large foreign inflows.

In other words, this is exactly the kind of environment where a headline-based risk gauge should show its strengths—and also reveal where headlines fail to pre-announce volatility.

This chart shows that the test window was far from a normal market regime. Instead of small, stable daily moves, the period featured repeated sharp swings in both directions, including several sessions that exceeded the 2% threshold I use to define a “big day.” That matters because a headline-based risk indicator becomes most informative in exactly this kind of environment: when volatility is large enough to reveal signal, but uneven enough to expose the model’s strengths and blind spots.

4) How I labeled regimes (operational thresholds)

From the distribution in this mini sample, ~1.7 behaves like a “typical day” region, while 2.0+ represents a clear step into higher uncertainty.

So operationally:

Calm: below ~1.7 → “tomorrow likely smaller move”
Alert: around ~1.7+ → “big move becomes more plausible”
Strong / High alert: ~2.0+ (especially ~2.5+) → “risk expansion mode”

This is not a universal constant—it’s a practical threshold that can (and should) be recalibrated as the dataset grows.

This plot helps justify the thresholds I use in practice. Most observations cluster in the mid-range, especially around the 1.6–1.8 area, which is why I treat that band as the “typical day” zone. By contrast, readings near 2.0 stand out as clear step-ups in headline intensity, while the spike above 2.5 represents a much stronger risk signal than the rest of the sample. In other words, these thresholds are not arbitrary—they reflect the actual shape of the distribution in this test window.

5) Accuracy, evaluated the right way: volatility first, direction second

I evaluate performance in two layers:

Layer 1 — Volatility: “Big day or not?”

I define a big day as a next-day move larger than 2% in absolute terms (up or down). Then I check whether News VIX was in an alert zone the day before.

Results (11 observations):

Overall big-day classification accuracy: 7 / 11 = 63.6%
When News VIX flagged risk, the next day was a big day 60% of the time (precision).
It also captured 60% of all big days (recall).
Confusion summary: 3 true positives, 2 false positives, 2 false negatives, 4 true negatives.

Interpretation: in this window, News VIX acted like a usable risk-regime switch—not perfect, but meaningfully informative for “tomorrow could be large; manage risk accordingly.”

Layer 2 — Direction: “If I had a tilt, was it right?”

Directional calls are inherently harder, so I only score direction on days where the model issued a non-neutral tilt.

In this window:

Non-neutral direction calls occurred 5 times (upward bias / strong upside / downside hint).
The tilt matched the next day’s sign 4 out of 5 times (80%).

Caveat: one of those “correct” calls was an almost flat day (+0.06%), so the more conservative interpretation is: the directional layer looks promising, but it needs a larger sample and stricter significance rules (e.g., don’t count near-zero days).

The scatter plot below reinforces the main point of this model: News VIX is more useful as a volatility filter than as a directional predictor. The relationship is not perfectly linear, and the sample is too small to claim a stable statistical rule. Still, the broad pattern is informative: higher News VIX readings tend to appear more often alongside larger next-day moves in absolute terms. In practice, that makes the indicator useful as a warning signal that tomorrow may fall outside a normal trading range, even when it cannot reliably predict the sign of the move.

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

(Full log is provided in the PDF table.)Here is the narrative review, focusing on what each signal implied and what the market actually did.

1/29 → 1/30 (News VIX 1.67; negative share ~20%)

Signal: Calm/borderline; mild constructive toneOutcome: +0.06% (flat)Read: Calm regime matched: low realized movement.

1/30 → 1/31 (News VIX 2.00; negative share ~19%)

Signal: Strong alert (expect risk expansion); tone looked upward-leaningOutcome: -5.26% (major down day)Read: Volatility call was correct (big move), direction was wrong.This is a classic VIX-type behavior: high uncertainty tells you “prepare for a large move,” but macro triggers and positioning decide the sign.

2/2 → 2/3 (News VIX 1.73; negative share ~21%)

Signal: Alert—big move possible; tone not deeply risk-offOutcome: +6.84% (massive rebound)Read: This is where the framework shines: elevated uncertainty preceded an outsized next-day move—here, a sharp rebound.

2/3 → 2/4 (News VIX 1.67; negative share ~29%)

Signal: Calm/borderline; tone mixedOutcome: +1.57% (moderate up)Read: Not a “big day” by definition; consistent with calmer regime.

2/4 → 2/5 (News VIX 1.00; negative share ~35%)

Signal: Calm by intensity (miss), but negative share was high → downside tilt warningOutcome: -3.86% (big down day)Read: Key lesson: sometimes intensity stays low, but negative concentration matters.This day supports the idea that tone can carry direction, even when “headline intensity” fails to anticipate magnitude.

2/5 → 2/6 (News VIX 2.00; negative share ~24%)

Signal: Strong alert (possible overreaction)Outcome: -1.44% (not big)Read: False alarm—cost of being cautious. When you set an alert threshold to catch storms, you’ll sometimes brace for one that doesn’t arrive.

2/6 → 2/7 (News VIX 1.47; negative share ~23%)

Signal: Calm (miss)Outcome: +4.10% (big up day)Read: Another blind spot: not all volatility is pre-announced by headline intensity. External catalysts, global spillover, and flow/positioning can dominate.

2/9 → 2/10 (News VIX 1.80; negative share ~23%)

Signal: Alert (over)Outcome: +0.07% (flat)Read: Another false alarm—suggests the model may need a “noise filter” for days when headlines are active but markets are already positioned.

2/10 → 2/11 (News VIX 1.33; negative share ~26%)

Signal: Calm; neutral toneOutcome: +1.00% (small)Read: Calm regime matched.

2/11 → 2/12 (News VIX 2.67; negative share ~6%)

Signal: High alert + strongly constructive toneOutcome: +3.13% (big up day; 5500 breakthrough narrative) Read: Ideal alignment: intensity flagged a big day, and tone correctly leaned upward.

2/12 → 2/13 (News VIX 1.60; negative share ~27%)

Signal: Calm; neutralOutcome: -0.28% (small)Read: Calm regime matched.

7) What this mini backtest suggests (professional takeaway)

What News VIX already does well

It functions as a risk-regime switch: when News VIX is elevated, the next day is more likely to produce outsized moves.
In a market like early Feb 2026—where KOSPI saw multiple 3–7% swings—this “big-move radar” is practically useful.

Where direction needs improvement

Direction is not stable from intensity alone.
Direction improves when headline tone becomes extreme (negative share spikes or positive dominance becomes clear), but macro shocks can override tone.

The cleanest mental model is:

News VIX = “how big”
Headline tone = “tilt”
External catalysts decide the final sign in shock regimes

This view makes the model’s main use case clearer. Once direction is removed, the key distinction is whether the next session stays within a normal range or breaks into a “big move” regime. That is the part of the forecasting problem where News VIX appears most practically useful.

8) News links (background reading for this window)

If you want to cross-check the macro context of each major swing: