REPLYOPSAI RESEARCH NOTE
EXPERIMENT · V6 · FORWARD TEST

V6 FreqAI PAPER Trading Experiment

An ongoing machine-learning experiment that retrains on incoming market data and trades BTC, ETH and SOL in simulated execution. The purpose is to observe the system forward—not to turn a small winning sample into a profitability claim.

PAPER — NOT LIVE MONEY

Current snapshot

Auto-updated from a sanitized metrics feed
+$10.42realized PnL5closed trades2open trades100.0%win rate*

*Win rate without sample size, payoff distribution and drawdown can be misleading.

FORWARD RECORD

Cumulative realized PAPER PnL

Closed trades only · not account equity
START $05 CLOSED TRADES+$10.42
EXPECTANCY / TRADE+$2.08PROFIT FACTOR∞*MAX CLOSED-PNL DRAWDOWN$0.00

Recent closed PAPER trades

MARKETSIDEEXITPAPER PNL
SOL09-24 17:40SHORTexit signal+$0.35
ETH09-24 21:32LONGroi+$2.00
ETH09-24 14:55SHORTexit signal+$2.28
SOL09-24 14:50LONGtrailing stop loss+$2.07
ETH09-24 14:32LONGexit signal+$3.72

What we are testing

V6 tests whether a retrained ML workflow can produce useful forward behavior after the entire pipeline is included: market data, features, model training, signal generation, position management and simulated execution. We are interested in stability and net outcomes, not just prediction accuracy.

MarketsBTC / ETH / SOL

Multiple crypto markets expose the system to overlapping market risk.

Timeframe5m

Signals and execution are evaluated on short-horizon market data.

Training window30 days

The model learns from a rolling recent sample rather than a fixed historical dataset.

RetrainingEvery 6h

New observations periodically produce a refreshed model.

Experiment pipeline

OKX DATA→FEATURES→FREQAI→SIGNAL→PAPER TRADE→METRICS

The public website receives only sanitized aggregate metrics. Exchange credentials and bot controls remain on separate trading infrastructure.

How we judge the experiment

Performance

Net realized PnL, payoff distribution, drawdown and losing sequences—not win rate in isolation.

Behavior

Trade frequency, simultaneous exposure, long/short behavior and changes across market conditions.

Learning stability

Whether periodic retraining improves, destabilizes or simply changes behavior.

Operations

Data freshness, process continuity and whether the deployed system matches the intended version.

What the current result does not prove

A small sample can produce extreme win rates by chance. PAPER fills do not reproduce all live slippage, latency and liquidity constraints. Future trades can reverse the current result, and retraining can change behavior. For those reasons, this page reports observations rather than extrapolating an expected return.

Why we keep V4 and V4.4 visible

A new model should not erase the systems that preceded it. Keeping earlier versions visible reduces survivorship bias and makes it harder to present only the experiment that currently looks best.

Compare V4, V4.4 and V6 →
Last metrics update: 2026-09-24T23:13:17.616Z

How this version was built

The build log explains why V6 followed the custom adaptive engine and how Freqtrade, FreqAI, OKX and the sanitized metrics pipeline fit together.

Read the V6 build log →