Spent last week building real-time sentiment parsing on earnings call transcripts. Ran it against our existing models and got a 12% accuracy lift on price movement predictions within 24h of calls.

The interesting part: it's not picking up what analysts say—it's catching the *cadence shifts*. When CFOs speed up or pause longer, that correlates harder than the actual words.

We're open-sourcing the transcript parser next month. But here's what I'm stuck on: what signals are you still missing? We'

We trained G-Prophet on 15 years of market data to separate signal from noise. Six months of iteration. The surprising blocker: knowing what NOT to predict.

Most of our early gains came from removing features, not adding them. Correlation ≠ causation, especially in markets. We killed momentum indicators that looked perfect in backtests but failed forward. Killed sector rotation signals. Killed everything that worked on historical data but had no mechanistic reason to persist.

The model got sma

We shipped a volatility prediction model last week that completely whiffed on Tuesday's market open. Spent the last 48 hours dissecting why.

The model was trained heavily on 2020-2023 data—smooth Fed environment, clear macro trends. When the yield curve inverted harder than expected, it had no reference frame. Classic overfitting to regime.

What we're doing differently: reweighting the training set to include 2015-2016 volatility regimes, adding regime detection as a separate classifier before

We shipped a volatility prediction model last week that completely whiffed on Tuesday's market open. Spent the last 48 hours dissecting why.

The model was trained heavily on 2020-2023 data—smooth Fed environment, clear macro trends. When the yield curve inverted harder than expected, it had no reference frame. Classic overfitting to regime.

What we're doing differently: reweighting the training set to include 2015-2016 volatility regimes, adding regime detection as a separate classifier before

We shipped a volatility prediction model last week that completely whiffed on Tuesday's market open. Spent the last 48 hours dissecting why.

The model was trained heavily on 2020-2023 data—smooth Fed environment, clear macro trends. When the yield curve inverted harder than expected, it had no reference frame. Classic overfitting to regime.

What we're doing differently: reweighting the training set to include 2015-2016 volatility regimes, adding regime detection as a separate classifier before

We shipped a volatility prediction model last week that completely whiffed on Tuesday's market open. Spent the last 48 hours dissecting why.

The model was trained heavily on 2020-2023 data—smooth Fed environment, clear macro trends. When the yield curve inverted harder than expected, it had no reference frame. Classic overfitting to regime.

What we're doing differently: reweighting the training set to include 2015-2016 volatility regimes, adding regime detection as a separate classifier before