Does FreqAI Work? Our Forward PAPER Trading Results
We are testing FreqAI on unseen market data in a continuously running PAPER trading environment. This page explains the published V6 results, including profitable and losing trades, and is updated as the experiment continues. The complete published trade record remains in the experiment log.
PAPER TRADING — NO REAL MONEY
This is an analytical companion to the primary V6 experiment log. It does not replace the source record or claim that FreqAI has been proven profitable. “Continuously running” describes the forward-testing approach, not a verified uptime guarantee.
Current Results
Version: V6 FreqAI. Status: Forward PAPER observation — fresh snapshot.
Mode: PAPER. Running since: not published in the source record. Experiment duration is not inferred from trade timestamps.
Last updated: . This is the source snapshot timestamp, not the page-request time.
Net PnL covers realized closed trades. Drawdown is the decline on the recorded cumulative closed-PnL path; it is not account-equity drawdown and does not measure unrealized risk in open positions.
What We Are Testing
V6 asks whether a periodically retrained machine-learning workflow can produce useful forward trading behavior after the full pipeline is included: market data, features, training, signals, position management and simulated execution.
The question is about net outcomes, stability and risk, not prediction accuracy alone. A model that predicts direction can still lose money when its signals become trades. The observed result belongs to this V6 implementation, not to every strategy that uses FreqAI.
Experimental Setup
The published V6 configuration and build log confirm the following setup. These are documented experiment settings, not a remote inspection of the trading server.
Public results come from the same sanitized snapshot used by V6 and Experiments. Exchange credentials and bot controls remain on separate infrastructure.
V6 architecture and configuration →Why Forward PAPER Testing?
A backtest evaluates a candidate on historical data. It is useful for investigating a hypothesis, checking implementation and testing sensitivity, provided the evaluation controls for leakage and repeated selection.
A forward PAPER test observes decisions as new market data arrives, rather than replaying a historical period selected in advance. Future observations are unavailable at decision time. Retraining can use newly observed history for subsequent decisions; that does not turn earlier trades into new predictions.
Forward observation can expose signal frequency, changing market conditions, position overlap and operational failures. It still requires checks for feature leakage and version discipline, and simulated fills do not establish live execution quality.
Backtest vs PAPER: evidence and limitations →Complete Results
The aggregate above covers the published V6 closed-trade record, including winning and losing outcomes. It is not a selected winning-trade subset. Open-position count is shown separately because unrealized PnL is not published in this snapshot.
Cumulative realized PAPER PnL
The chart uses the source PnL path, ordered by recorded closed trades. It is not a price chart or a time-normalized return series.
Recorded entries: 20. Wins: 16. Losses: 3. Win/loss categories are reported as supplied, without inferring missing classifications.
Mean realized PnL per closed trade (source expectancy): +$1.08.
Profit factor: 2.947. This ratio is unstable in a small sample or when few losing trades have been observed.
“Complete” refers to the published record, not every dimension of risk. The public feed does not provide starting capital, a full mark-to-market equity history or a cost breakdown, so this page does not supply percentage returns, annualized performance or reconstructed live-equivalent results.
Trade Record
The shared trade journal contains published open/close timestamps, prices, directions, exit reasons and realized PAPER PnL. Select V6 FreqAI to inspect its record, including losing trades. The primary V6 page also shows recent trades.
What Would Invalidate the Experiment?
A formal pre-registered failure threshold was not established in the published V6 methodology. This analysis does not introduce a new trading rule, drawdown cutoff, minimum trade count or automatic stop condition.
The existing methodology identifies the following ways the hypothesis can fail or its evidence can become unreliable. They are evaluation dimensions, not retrospective numerical pass/fail thresholds:
- Economic failure: forward net outcomes fail to support a useful edge, or losing sequences and drawdown undermine an attractive headline win rate.
- Behavioral or learning failure: retraining destabilizes decisions, position overlap creates conflicting exposure, or prediction quality fails to translate into useful trades.
- Evidence failure: future information leaks into features or evaluation, materially changed versions are pooled, or results cannot be attributed to the deployed V6 record.
- Operational failure: missing or stale data, interrupted processes or a mismatch between the intended and deployed configuration prevent a reliable interpretation of the sample.
A losing trade alone does not invalidate a strategy. A stale snapshot means current evidence is unavailable; it is not proof that the trading hypothesis has failed. Decisions should preserve the full record and state their rationale.
Published methodology →Limitations
- PAPER trading ≠ live execution. No real money is represented; simulated fills do not establish achievable live returns.
- Sample size. There are 19 closed trades in the current snapshot. This is a limited sample and is insufficient for a statistically strong profitability conclusion.
- Fees and slippage. We report the source field labeled net PnL without changing its cost assumptions. The public snapshot does not include a fee schedule, per-trade fee breakdown or slippage model, so their completeness cannot be independently verified here. We do not deduct guessed costs or assert that live fees, liquidity, latency and slippage are fully reproduced.
- Market regime dependence. BTC, ETH and SOL can share market risk. Results from one observed period do not establish robustness in different volatility or trend conditions.
- Measured risk is incomplete. Closed-PnL drawdown excludes unrealized open-position losses. Missing start dates and capital denominators prevent defensible duration-normalized or capital-normalized comparisons.
- Version and model dependence. Retraining changes the model. The record evaluates the V6 workflow and should not be treated as the performance of an unchanged model across all trades.
- No future-profitability inference. Past or current PAPER performance does not establish future profitability or justify a claim that FreqAI works in general.
Current Conclusion
The recorded sample has 19 closed PAPER trades, 16 wins and 3 losses, with positive realized net PnL. This describes a positive observed sample, not proof of a durable trading edge. The sample is small and insufficient for a statistically strong conclusion. Current PAPER performance does not establish future or LIVE profitability.
Read net PnL alongside trade count, observed losses and closed-PnL drawdown. A high win rate cannot settle the question without payoff size, costs, risk and a defensible evaluation procedure.
Update History
— This analytical research page was created and first published, with the V6 experiment log retained as the primary evidence source.
Latest source refresh: . Metrics and the current conclusion update from the source snapshot on page load. Automated data refreshes are distinct from editorial revisions and do not imply a new strategy version.
No earlier publication dates or historical research updates are reconstructed. Future material methodology or interpretation changes should be recorded here explicitly.