MACHINE LEARNING / FREQAI

Prediction is not
trading evidence.

A model can score well and still produce a poor trading system. We focus on the complete path from training data to forward PAPER outcomes.

MLMODEL
FEATURESTARGETRETRAINSIGNAL
THE PIPELINE

Where ML experiments break.

Leakage, unstable targets and regime change can make predictive metrics look stronger than forward trading evidence.

01FeaturesMarket observations
→
02TrainingRolling window
→
03PredictionModel output
→
04TradeNet result
LEARNING PATH

Build the experiment in the right order.

01 / START

FreqAI setup

Define the experiment, isolate PAPER execution and get a first model running without confusing setup success with profitability.

Start here →
02 / TRAIN

Training and leakage

Windows, cadence and future information.

03 / MODEL

Choosing models

Complexity versus useful trading evidence.

04 / EVIDENCE

Is FreqAI profitable?

What can actually be concluded.

EVALUATION STACK

Measure the model and the trading system separately.

MODEL

Prediction quality

Validation metrics tell us how a model behaves against its target, not whether trades are profitable.

POLICY

Decision quality

Thresholds, entries, exits and sizing determine how model output becomes market exposure.

RESULT

Net forward evidence

Closed-trade expectancy, drawdown, costs and stability on new data determine what the complete experiment actually produced.

FORWARD EXPERIMENT

Watch V6 learn on new data.

The current V6 FreqAI branch retrains periodically and records its PAPER trades without rewriting earlier system histories.

View current results
TRAINING30dRETRAIN6hMARKETS3