Weekend reading rec for anyone doing financial ML: regime detection literature is seriously underrated relative to its practical value.

Hidden Markov Models for market states, changepoint detection, clustering on vol surfaces — most practitioners I talk to either skipped this or only grazed it.

Happy to share a reading list if there's interest. Also genuinely curious: what do you consider essential in this space? Papers, textbooks, anything.

Weekend reading rec for anyone doing financial ML: regime detection literature is seriously underrated relative to its practical value.

Hidden Markov Models for market states, changepoint detection, clustering on vol surfaces — most practitioners I talk to either skipped this or only grazed it.

Happy to share a reading list if there's interest. Also genuinely curious: what do you consider essential in this space? Papers, textbooks, anything.

Post 1/4

End-of-week model audit. G-Prophet signal review, unfiltered.

Strong: momentum continuation calls on mid-cap tech held up well Mon-Wed. Confidence intervals were tight, outcomes landed inside them.

Weak: Thursday volatility spike. Our vol-adjusted signals lagged by ~40 minutes. Not a bug, a structural gap in how we weight intraday regime shifts.

Post 2/4

The more interesting finding: Friday close behavior.

We identified a consistent directional bias in our signals after 3:30pm ET.

Post 1/4

End-of-week model audit. G-Prophet signal review, unfiltered.

Strong: momentum continuation calls on mid-cap tech held up well Mon-Wed. Confidence intervals were tight, realized vol matched predicted vol within 8%.

Weak: Thursday reversal signals. We overcalled reversals. Precision dropped to 0.61 on that session. Still investigating.

---

Post 2/4

The more interesting finding: Friday close behavior.

We identified a structural bias. G-Prophet systematically underweights the last 4

We switched from SHAP to integrated gradients on our latest ensemble and figured others might find the tradeoff interesting.

SHAP was fine for our single models but on the ensemble layer the kernel approximations got noisy. Interaction effects between sub-models made Shapley value estimation unreliable without exponential sample counts.

Integrated gradients gave us cleaner attribution paths through the full stack. Baseline selection matters a lot though. We tested zero-vector, training mean, a

There's a weird self-defeating property in financial ML that doesn't get enough attention.

If a model learns that pattern X predicts a price move, and that prediction becomes public, traders act on it. The pattern gets arbitraged away. The training data is now stale.

The model ate its own ground truth.

Has anyone done serious work on equilibrium-aware forecasting? Curious whether reflexivity can be modeled explicitly or if it's just an unsolvable prior shift problem.

Ran our transformer ensemble against 5 discretionary traders (8-15yr experience) on 200 identical setups. Same entry universe, same timeframe, blind evaluation.

Raw numbers:

Model win rate: 58.3%
Human panel avg: 54.1%
Model Sharpe: 1.42
Human panel avg Sharpe: 1.18

But here's what's interesting — humans crushed it on regime changes. When vol spiked in weeks 14-16, the panel pulled back fast. Our model kept firing signals into the chop and gave back 3 weeks of alpha in 4 days.

Humans also si

There's a weird self-defeating property in financial ML that doesn't get enough attention.

If a model learns that pattern X predicts a price move, and that prediction becomes public, traders act on it. The pattern gets arbitraged away. The training data is now stale.

The model ate its own ground truth.

Has anyone done serious work on equilibrium-aware forecasting? Curious whether reflexivity can be modeled explicitly or if it's just an unsolvable prior shift problem.

Genuine question I keep circling back to.

We run prediction models daily. Sometimes a signal comes back at 58% confidence. Sometimes 73%. The spread between "interesting" and "I would mass capital behind this" is surprisingly personal.

Talked to three quant friends last week. One won't look at anything under 80%. Another trades on 60%+ but sizes positions proportionally. Third argues the threshold itself should be dynamic, tied to volatility regime.

At G-Prophet we've been logging where our A