Updated November 2026 · First published March 26, 2026

Why GodFin Stopped Predicting Prices

GodFin started out building LSTM models to forecast price direction. The first version of this post explained how we tried to correct their bias. In July 2026 we audited those models, found no measurable predictive edge, and permanently retired them. This is the rewritten version: what went wrong, and why the part we kept is a set of hard risk limits.

The failure mode we kept fighting

A sequence model trained on price history can look healthy on a validation chart and still lean the same way, bullish or bearish, long after the market has changed. We saw this over and over. The reasons are well known:

  • Training data contains regime imbalance. A long bull market teaches optimism.
  • Loss functions reward average directional accuracy more than calibration.
  • Feature distributions drift, so the model's internal baseline goes stale.
  • Trading decisions are discrete. A small skew in the output pushes many names over the same threshold at once.

The model does not look broken. It just keeps preferring one side of the market, and a portfolio built on it becomes structurally long or short without anyone deciding that it should.

Prediction bias is dangerous because it masquerades as conviction. The model is not seeing more edge. It is carrying a stale baseline forward.

What we tried

We normalized the raw output against its own recent history with a rolling z-score, so the question became "is this output unusual for this model?" instead of "is it positive?" We added confidence estimates and blended in sentiment. Each step made the output look cleaner and more consistent across tickers.

None of it created an edge. When we audited the models against what actually happened, we found no measurable predictive edge. Cleaner-looking output had made us more confident in it, which is the same failure mode in a new form.

Why we retired the models instead of tuning them

We could have kept tuning. But a product that sells forecasts has a built-in problem: the more convincing the forecast looks, the bigger the position people take on it. A biased model that looks highly confident is the worst case, because it invites the largest trades at the moment it is most likely to be wrong.

So in July 2026 we shut the prediction models down for good. GodFin makes no market predictions and sends no trading signals.

What we kept: limits that do not care how sure you are

The part of the old system that held up was the part that ignored the forecast: the risk rules. GodFin is now a pre-trade risk gate. You place the order, and the gate checks it against three hard limits before it reaches your broker:

  • 2% max risk per trade. The order is sized so that if the stop is hit, you lose at most 2% of equity.
  • 5% daily-loss lockout. Once your account is down 5% on the day, new risk is blocked for the rest of the trading session.
  • 20% drawdown kill switch. While your equity is 20% or more below its peak, new risk is blocked.

Orders that close or reduce a position always pass, so you can always get out. None of the three rules asks where the idea came from. A model, a newsletter or your own gut gets exactly the same treatment, and that is the point.

The lesson

Conviction is cheap to produce and expensive to act on. A forecast can be wrong in ways you will not see until after the trade. A limit on how much any single trade, day or drawdown can cost you is right regardless of which way the market goes. That is the product we decided to build.

Join the GodFin waitlist

Get early access to the risk gate: 2% per trade, a 5% daily-loss lockout and a 20% drawdown kill switch on every order. No predictions, no signals.

Join the waitlist