The scoreboard

Every model behind a verdict was tested on years it never saw, against a simple baseline anyone could compute. When a model loses, this page says so and the app stops using it.

Expanding-window walk-forward by calendar year. Training data stops 25 trading days before each test year so no target overlaps it. Every number below is out-of-sample. Test years 2019–2026, Nifty 50 stocks, data to 14 Oct 2026.

Summary

Return range: +1.0% better than the baseline, but its 80% range only caught 75.1% of outcomes.
Marginal
Up-or-down calls: right 55.2% of the time. Just assuming "up" was right 57.4%. So the app never makes direction calls.
Loses
Volatility: GARCH(1,1) cuts forecast error by 28.5%. The LSTM (deep learning) manages 24.8%, good, but the 1986 model still wins.
Beats baseline
Market regime: days called "Stressed" were followed by 1.9× the volatility of "Calm" days.
Informative
Earnings-call language drift: Needs a scraped transcript archive first. No numbers are shown until it has been backtested like the others.
Not built

20-day return range

Marginal

Model: LightGBM quantile regression (10th / 50th / 90th pct)
Baseline: The stock's own trailing 2-year spread of 20-day returns
Scored by: Pinball loss (lower is better); skill = 1 - model/baseline

Measure	Model	Baseline
Pinball loss skill vs baseline	+1.0%	—
10–90% range caught the outcome	75.1%	78.4%
Direction right	55.2%	57.4% (always "up")

model · baseline (drawn side by side at each percentile). On the dashed line means the quantiles are honest: the 10th percentile should be undercut 10% of the time.

Skill by test year (right of the line = beats baseline)

2019+0.9%
2020+0.5%
2021+0.8%
2022+0.2%
2023+0.6%
2024+2.2%
2025+4.5%
2026−1.3%

20-day volatility

Beats baseline

Model: LSTM (deep learning) vs GARCH(1,1)
Baseline: Last 20 days' volatility carried forward
Scored by: QLIKE loss (lower is better); skill = 1 - model/baseline

Contender	QLIKE loss	Skill vs baseline
LSTM	0.331	+24.8%
GARCH(1,1)	0.315	+28.5%
Trailing 20-day vol	0.441	—

LSTM skill by year (right of the line = beats baseline)

2019+36.7%
2020−7.5%
2021+28.7%
2022+36.4%
2023+22.9%
2024+38.9%
2025+36.6%
2026+45.8%

GARCH skill by year (right of the line = beats baseline)

2019+37.6%
2020+10.8%
2021+38.5%
2022+28.7%
2023+10.9%
2024+48.0%
2025+26.7%
2026+48.5%

Market regime

Informative

Model: 3-state Gaussian HMM on Nifty 50 returns and volatility, forward-filtered
Baseline: No regime: every day treated alike
Scored by: Does the regime called today separate the NEXT 20 days? (out-of-sample)

Regime called	Days (test years)	Next 20 days: volatility	Next 20 days: Nifty return
Calm	805	11.1%	+0.9%
Choppy	700	15.2%	−0.0%
Stressed	395	21.1%	+2.0%

The regime separates future volatility cleanly. It does not reliably separate future returns, so the app uses it only for claims about the market backdrop.

Earnings-call language drift

Not built

Model: FinBERT + embedding drift between consecutive concalls
Baseline: None yet
Scored by: None yet

Needs a scraped transcript archive first. No numbers are shown until it has been backtested like the others. The plan: scrape BSE concall PDFs, score each call with FinBERT, measure how far its language drifts from the previous quarter, then test whether drift predicts the next quarter's surprise.

Want to see what these models say about your view?

Test a thesis