UNSUPERVISED · KMEANS · 5,711 DAYS · 2004–NOW
Market States
The machine grouped 22 years of trading days into five regimes — no labels given, just the shape of VIX, the yield curve, and S&P momentum & turbulence. Below: which regime today sits in, and the days that looked most like it.
TODAY · 2026-09-25
Range-bound calm
VIX
15.26
10y-2y
0.31pp
S&P 20d
-0.1%
S&P vol
10.62%
WHEN THE MARKET LAST LOOKED LIKE THIS — WHAT FOLLOWED
AVG +30 DAYS
-1%
AVG +90 DAYS
+2%
2025-02-20Range-bound calmVIX 15.7-11.8%+1.3%
2006-05-12Range-bound calmVIX 14.2-3.2%+2.6%
2019-05-10Range-bound calmVIX 16+2.2%+4.3%
2018-06-21Range-bound calmVIX 14.6+3.3%-3.9%
2019-03-25Range-bound calmVIX 16.3+3.1%+5.5%
2007-06-11Range-bound calmVIX 14.7+0.1%+2.1%
Nearest neighbours in standardized feature space (the last 60 days excluded so matches aren't trivially recent). +30 / +90 columns are what the S&P 500 actually did after each date — history, not a forecast. Small sample; the future is not obliged to rhyme.
DOES THE SIGNAL HAVE AN EDGE? · S&P 500 forward return, 22 years
BASELINE · ANY DAY+3.5%72.8% hit · n=5621over the next 90 trading days, the bars below show distance from this
BY MARKET REGIME
Volatile recovery
+11.46%+7.96pp89.7% hit+8.83%…+16.35%n=273
Crash / capitulation
+4.61%+1.11pp51% hit-16.41%…+26.22%n=102
Range-bound calm● NOW
+3.49%-0.01pp≈ baseline · no edge+0.13%…+7.92%n=2487
Correction / stress
+3.14%-0.36pp≈ baseline · no edge-3.02%…+11.97%n=834
Steady climb (steep curve)
+2.49%-1.01pp70.4% hit-1.11%…+7.19%n=1925
BY VIX (FEAR)
< 15 · complacent
+2.75%-0.75pp75% hit0%…+6.45%n=2092
15–20 · normal● NOW
+2.98%-0.52pp≈ baseline · no edge-0.87%…+8.06%n=1754
20–30 · stressed
+3.59%+0.09pp≈ baseline · no edge-2.3%…+11.41%n=1326
≥ 30 · panic
+8.81%+5.31pp76.8% hit+1.47%…+18.54%n=449
BY YIELD CURVE (10Y–2Y)
inverted · < 0
+6.5%+3pp87.5% hit+2.93%…+9.73%n=786
flat · 0–0.5● NOW
+1.92%-1.58pp61.1% hit-2.98%…+6.9%n=1176
normal · 0.5–1.5
+3.02%-0.48pp≈ baseline · no edge+0.2%…+8.05%n=1879
steep · ≥ 1.5
+3.74%+0.24pp≈ baseline · no edge-1.17%…+9.3%n=1780
Forward returns are what actually happened, not a prediction. · Where a bucket's returns ≈ the baseline, the signal carries no usable edge. · Small samples (n < 30) are noisy — read the dispersion (p25–p75), not just the mean. · Past regimes are not obliged to rhyme with the future.
DOES THE TRADE TAB'S SIGNAL HAVE AN EDGE? · SET INDEX · 26 YEARS
What the conventional retail playbook says (RSI<30 = buy, lower BB = buy, pullback = buy) vs what 26 years of SET history actually shows. The bars are centred on the baseline — a bar at centre means no usable edge on this history.
BASELINE · ANY DAY (SET)+2.9%56.8% hit · n=6359over the next 90 trading days
RSI(14) — is oversold/overbought predictive on SET?
The finding: SET rewards momentum continuation, not mean reversion.
< 30 · oversold
+2.51%-0.39pp≈ baseline-5.04…+10.8%n=752
30–45 · low
+1.12%-1.78ppmild drag-6.4…+7.78%n=1520
45–60 · neutral● NOW
+2.55%-0.35pp≈ baseline-5.76…+9.95%n=1759
60–70 · high
+3.3%+0.4pp≈ baseline-5.36…+9.85%n=1058
> 70 · overbought
+5.44%+2.54ppedge ↑-3.76…+12.39%n=1270
EMA(9/21) CROSS — trend confirmation
EMA9 > EMA21 · bull● NOW
+3.97%+1.07ppmild edge-4.97…+10.69%n=3510
EMA9 < EMA21 · bear
+1.59%-1.31ppmild drag-5.78…+9.63%n=2849
BOLLINGER BAND POSITION (%B) — does 'touch the band' work?
Below lower band ≠ buy signal on SET history.
below lower band · %B < 0
+0.8%-2.1ppnegative ↓-6.57…+8.63%n=358
lower half · 0–0.5● NOW
+1.94%-0.96ppmild drag-5.69…+9.18%n=2475
upper half · 0.5–1
+3.72%+0.82ppmild edge-5.17…+11.05%n=3214
above upper band · %B > 1
+4.64%+1.74ppmild edge-2.98…+11.32%n=312
SET 20-DAY MOMENTUM — buy dips or ride momentum?
Sharp pullbacks show no edge. Strong rallies have the largest edge in this dataset.
< -8% · sharp pullback
+0.42%-2.48ppnegative ↓-10.84…+14.54%n=379
-8 to 0% · negative
+2.01%-0.89ppmild drag-5.57…+8.25%n=2512
0–8% · positive● NOW
+3.11%+0.21pp≈ baseline-5.04…+10.08%n=2968
> 8% · strong rally
+8.1%+5.2ppedge ↑-4.34…+20.68%n=500
Forward returns are SET actual history, not a prediction. · RSI and BB signals are computed on the SET index, not individual stocks. · Where a bucket ≈ baseline it carries no usable edge on this history. · Small samples (n<30) are noisy — read dispersion (p25–p75), not just the mean. · The past need not rhyme with the future.
WHAT STATISTICALLY PRECEDES THE SET? · GRANGER CAUSALITY · 0 DAYS
Granger causality is not a trading signal. It means that past values of a series helped predict SET returns in a linear regression over the historical sample — nothing more. Relationships break, markets adapt, and statistical significance in a backtest does not guarantee a live edge. Treat these as structural context for understanding what moves the SET, not as reasons to buy or sell.
STATISTICALLY SIGNIFICANT (p<0.05)
None at p<0.05 in this sample.
NOT SIGNIFICANT — shown for completeness
Granger = predictive precedence, not true causation. ADF-tested for stationarity and aligned by trading session. · run 2026-09-21◐CACHED · —
ANOMALY WATCH · PRICE VS FUNDAMENTALS
IsolationForest over 25 SET names clustered on 14 fundamentals+price features. A flag means the stock trades in a way its cluster's numbers can't explain — the market is telling a story the statements don't carry yet. It is a pointer for looking, never a buy/sell call.
NAMECLUSTERANOMALY 0–120D RET
DELTA.BKgrowth_premium1.00-8.8%
PTTGC.BKfair_value0.59+26.3%
ADVANC.BKspeculative0.55-2.5%
BGRIM.BKfair_value0.470.0%
mined 2026-09-20T18:01:54.413556 · refreshed weekly by ingestion/pattern_mine.py
THE FIVE REGIMES · share of 22 years
Range-bound calm← TODAY45.1%
VIX ~15.22 · 10y-2y 0.26pp · S&P 20d +1.96% · vol 11.44%
Steady climb (steep curve)33.7%
VIX ~16.72 · 10y-2y 1.91pp · S&P 20d +1.62% · vol 12.31%
Correction / stress14.6%
VIX ~26.6 · 10y-2y 0.87pp · S&P 20d -4.68% · vol 23.08%
Volatile recovery4.8%
VIX ~32.47 · 10y-2y 1.57pp · S&P 20d +5.7% · vol 32.64%
Crash / capitulation1.8%
VIX ~57.02 · 10y-2y 1.7pp · S&P 20d -13.74% · vol 68.37%
LAST 252 TRADING DAYS · regime tape
Volatile recoveryRange-bound calmCrash / capitulationCorrection / stressSteady climb (steep curve)
KMeans (k=5) on z-scored FRED daily features; analogue = nearest neighbour in feature space, forward S&P 500 return reported · generated 2026-09-25. Credit spreads are excluded from the state model (FRED truncates the free series) but stay live in the Mr. Market gauge. See Data Integrity.