Open quantitative record
Bitcoin AI Trading Model – Forward Trades on Telegram, Results & Research
I build and run machine-learning Bitcoin trading models and publish their work in the open: forward trades on Telegram as they happen, the methodology and research behind each model, and the experiments that failed. The current production model is the Frontier Model, the exact frozen T1071 XGBoost BTCUSD daily signal model with a separately audited execution layer. Repeated rolling refits did not reproduce the persistence of the frozen binary, so production does not automatically retrain it. You can use the site either to learn from the process and develop your own Bitcoin trading system, or simply follow the model’s published forward trades as they unfold. Everything is free, with nothing to sell — no subscriptions, no broker affiliates and no copy-trading.
Live feed
Live Frontier Model BTC Trades
Frontier, our current production system, trades BTCUSD from daily candles and can publish LONG or SHORT signals. Across the complete January 2018 to August 2026 research track it produced 361 trades. In the strict model-weight OOS period from 30 July 2022 to 12 August 2026 it produced 210 trades — about 4.3 trades per month — with a median logged entry-to-exit timestamp difference of 96 hours. Because the historical research uses D1 bars, the exact intraday moment when a stop loss or take profit was hit is not available from the backtest. This is not a high-frequency system, and quiet periods between signals are normal.
Posts go out on Telegram when a position opens and again when its protection changes or the position closes. A new-trade message includes direction, entry and the stop loss immediately. Under the current Frontier execution rule, the strategy take profit is intentionally not active on the entry D1; it is attached from the start of the second D1, and that TP activation or any later TP change is published as a separate update. Closing messages identify take-profit or stop-loss executions. The same feed is mirrored below if you would rather not join the channel.
Record
Frontier Bitcoin backtest, In-Sample and OOS Results
Almost every published trading curve hides where a model was fitted and where its frozen weights were later exposed to unseen chronology. This one does not. Frontier keeps the research history split explicitly between the 2018–July 2022 in-sample / safe training block and the strict model-weight OOS period beginning with the first post-weight entry on 30 July 2022.
The frozen Frontier research track runs from January 2018 to August 2026 and contains 361 exact BTCUSD trades. The cumulative result is +987.6 simple percentage points, summed trade by trade: +733.7 points in the 2018–July 2022 in-sample block and +253.9 points in strict model-weight OOS. Over the 8.61-year track, that is approximately +114.7% per year on a simple annualised basis. That is a research statistic — it is not the same thing as saying that a fixed-size account returned +987.6% or earned a compounded CAGR of 114.7%.
The study includes 0.045% commission per side. Swap, funding and additional execution slippage are excluded from this historical research track, so real or simulated account results can differ materially.
Research leverage: 1:1 equivalent.
All historical return, drawdown and SQN figures shown below are unlevered and correspond to
1:1 notional exposure. They are not multiplied by the live sizing leverage.
The 2:1 figure shown in the table is the maximum live leverage ceiling,
not the leverage used to produce the research statistics.
In-sample and model-weight OOS — shown separately
| Metric | Model-weight OOS | In-sample |
|---|---|---|
| Annualised return (simple) | +62.9 pt/yr | +160.4 pt/yr |
| Total return | +253.9 pt | +733.7 pt |
| Win rate | 50.5% | 63.6% |
| Trades | 210 | 151 |
| Period covered | 30 Jul 2022 – 12 Aug 2026 | Jan 2018 – 29 Jul 2022 |
| Average trades per month | 4.3 | 2.8 |
| Median entry-to-exit timestamp* | 96 h | — |
| Average entry-to-exit timestamp* | 146.7 h | — |
| Maximum drawdown | 25.34 pt | 31.04 pt |
| Research leverage | 1:1 equivalent | 1:1 equivalent |
| Maximum live leverage ceiling | 2:1 | 2:1 |
*Annualised return is shown as a simple annualised research rate: total percentage points divided by calendar years in the corresponding block. Model-weight OOS: +253.9 points over 4.04 years = +62.9 pt/yr. In-sample: +733.7 points over 4.58 years = +160.4 pt/yr. This is not a compounded account CAGR. Historical statistics are shown at 1:1 notional-equivalent exposure; the 2:1 value is only the maximum live leverage ceiling. *Holding-time figures are differences between the logged D1 entry and exit timestamps. The OOS median is 96 h and the average is 146.7 h; the corresponding in-sample timing figures are not quoted here because they were not part of the final published phase summary. These D1 timestamps do not identify the exact intraday time at which SL or TP was touched. In live execution, the catastrophic SL is active from entry, while the strategy TP is attached from the start of the second D1. Full-track historical drawdown is approximately 31.04 points; strict model-weight OOS drawdown is 25.34 points. Past performance, simulated or real, is not a reliable indicator of future results.
Frozen-model report: validation summary, drawdowns, trade distribution and the current Frontier research record.
Independently verified account
The backtest above is a model study. The account below is the live-tracked record, read directly by FXBlue — a third party with no reason to flatter it. Treat the two as separate things, because they are.
Method
What is actually under the hood
“AI-powered” means very little on its own, so here is the specific version. Frontier Model is an XGBoost trading strategy — gradient-boosted trees classifying directional edge on daily Bitcoin candles. The production signal path is the exact frozen T1071 model using seven engineered inputs and no LSTM inference. Recurrent models were tested extensively earlier in the research programme and did not survive the same validation standard.
Features
Seven engineered inputs derived from price structure, realised volatility and normalized momentum/range behaviour. No discretionary override is required to generate a signal.
Validation
The current signal binary is deliberately frozen. Rolling retrain/refit tests were run across chronological windows and seeds, but they did not reproduce the persistence of the exact original T1071 weights. The final execution layer was then audited separately, one change at a time, without retraining the signal model. For further methodological context, see our research on overfitting in trading strategies and data leakage in financial machine learning .
Risk layer
Historical research statistics are reported at 1:1 notional-equivalent exposure. The public maximum live leverage ceiling is 2:1. The current execution geometry uses a catastrophic SL of 7.75 ATR from entry, TP 2.40 ATR for SHORT and 2.52 ATR for LONG, with the strategy TP activated from the start of the second D1.
What is not claimed
No price prediction and no forecasting of tops or bottoms. The model estimates conditional directional edge over a fixed horizon; the execution and risk layers decide how that signal is traded.
The current Frontier is the exact frozen T1071 XGBoost signal binary, not the later eight-input retrained candidate previously shown on this page. The recent audit first reproduced the frozen model exactly, then tested retraining frequency and exit changes separately. Rolling refits were weaker, while the frozen binary retained the stronger chronological persistence. The final public execution layer keeps the signal model unchanged and uses TP 2.40 ATR for SHORT, TP 2.52 ATR for LONG and a 7.75 ATR catastrophic stop, with SL active from entry and TP from D1 #2. The full write-up — feature construction, XGBoost, walk-forward analysis, out-of-sample chronology and the research graveyard — remains on the methodology page. The verified forward record is tracked independently on FXBlue. The research archive is available on GitHub.
What to expect
The uncomfortable parts, stated in advance
Most trading systems are abandoned during the flat stretch, not the losing one. Setting expectations before that happens is more useful than any performance chart.
Quiet periods are normal
The strict model-weight OOS period averages about 4.3 positions per month, while the complete 2018–August 2026 track averages about 3.5. Activity is not evenly distributed, and some stretches are much quieter.
Drawdowns are structural
A Bitcoin trading strategy with positive historical expectancy can still lose for extended stretches. The published drawdown is a historical research figure, not a future ceiling.
A roughly 50% OOS win rate is part of the strategy
Frontier Model won 50.5% of the 210 strict model-weight OOS trades and 63.6% of the 151 in-sample trades. The result depends on the size and path of wins and losses, not on win rate alone.
The system evolves
Frontier Model is the current production model and its signal binary is frozen. Material execution or model changes are dated and documented so the forward record remains interpretable rather than silently mixing different systems.
What you get here
Every trade, live and timestamped. New BTCUSD signals are posted to the free Telegram channel at execution — before the outcome is known. The feed also publishes the D1 #2 take-profit activation or later TP changes, and identifies TP or SL execution when the position closes. The tracked account remains independently visible through FXBlue.
The complete research record. The full methodology, out-of-sample results side by side with training data, walk-forward validation and rejected experiments — failures included.
Research you can build on. Regular articles on overfitting, walk-forward validation, XGBoost trading strategies and machine-learning Bitcoin trading models, written so you can develop and test your own.
If it is useful, it is yours.
Disclosure
Research curve and tracked execution are separate records
The Frontier Model backtest above is a historical model study. The FXBlue account is a separately tracked execution record. They are shown together for transparency, but they are not the same dataset and should not be read as if one proves the other.
The Frontier Model historical study includes 0.045% commission per side. Swap, funding and additional execution slippage are excluded from that research curve, so a live or simulated account can differ materially because of spread, fills, financing costs, venue rules and market stress.
The forward record is timestamped, the methodology is open, and the existing FXBlue account is read directly by a third party. That does not eliminate execution differences or future model risk; it simply makes the public record easier to audit.
Past performance, whether backtested, simulated or real, is not a reliable indicator of future results.
Bitcoin trading research
- XGBoost for Trading: Why Gradient Boosting Beats Deep Learning on Price DataOur companion article explains why the LSTM branch lost. This one explains why the production path is now XGBoost for trading: not because gradient boosting is fashionable, but because a … Read more
- Why Our LSTM Bitcoin Model Failed (and What Beat It)There are thousands of tutorials showing an LSTM Bitcoin prediction model with a beautiful chart of predicted versus actual price. Almost none of them show a trade net of fees, … Read more
- Walk-Forward Analysis for Bitcoin Trading: A Complete Worked ExampleAlmost every article about walk-forward analysis for Bitcoin explains the diagram and stops there: train here, test there, move the cutoff forward. The part that actually matters is harder — … Read more
Who writes this
Adrian Velai
Adrian Velai is a pen name. I keep my legal identity private, which is a deliberate choice and one I would rather declare than disguise — you should weigh it accordingly.
What I put in its place is the only thing that can actually be checked: a complete trade record independently tracked by FXBlue, the methodology written out in full, and record published openly. An anonymous author with an auditable record is a better bet than a named one without.
Questions
Common questions
Is this really free?
Yes. There is no paid tier, no private signal channel and no affiliate link to a broker. Running the site costs money, so there is a bitcoin address at the bottom of this page for anyone who wants to contribute. Nothing is gated behind it.
Why publish a working model instead of keeping it private?
A low-frequency Bitcoin trading model is not degraded simply by people knowing that it exists. In return, publishing gives the project a public, timestamped forward record — which is more useful than a private backtest nobody can audit.
Why is the account simulated?
It lets me publish a complete, independently tracked execution record without presenting a selectively edited real-money account. The trade-off is stated openly: simulated execution cannot reproduce every live fill, margin event, spread expansion or financing cost.
Can I follow the trades automatically?
There is no public copier, no downloadable trading EA and no managed-account service. The free Bitcoin signals are published for reading on Telegram. Any decision to act on them with an auto reading bot is yours alone.
What is the minimum account size?
That depends on your own circumstances, sizing and venue, so this site does not prescribe one. Historical research statistics are shown at 1:1 notional-equivalent exposure, while the public maximum live leverage ceiling is 2:1.
Which venues can Frontier be executed on?
Frontier generates BTCUSD trades rather than requiring one specific broker or exchange. The trade itself is venue-neutral, but execution costs are not.
For crypto exposure, I generally prefer a reputable exchange offering BTC perpetual futures over a crypto CFD when both are legally available to the user. The main reason is financing transparency. A CFD can apply an overnight swap or financing charge set by the broker, and that charge can materially affect positions held for several days. With a perpetual contract, the equivalent carrying cost is the published funding rate, which can be either paid or received depending on market conditions.
That does not mean perpetual futures are automatically cheaper. Funding can become expensive, spreads and commissions still matter, and leverage introduces liquidation risk. The advantage is that the financing mechanism is generally easier to observe, quantify and include in an execution audit.
The historical BitcoinAI.Pro research curve does not assume that one venue is universally best. Anyone evaluating the published signals should compare the total cost of execution — commission, spread, slippage, overnight swap or funding, and margin rules — at the venue they actually use.
How often does it trade?
The strict OOS period averaged about 4.3 trades per month. Its median logged entry-to-exit timestamp difference is 96 hours and its average is 146.7 hours. Because the historical research uses D1 bars, those are timestamp-resolution statistics rather than exact intraday holding times. In live execution the catastrophic SL is active from entry and the strategy TP is attached from the start of D1 #2; either can then close the trade intraday when its level is reached. Activity is not uniform, so some weeks can still be quiet.
Do you offer advice or manage accounts?
No. I do not manage money, hold client funds, give personal recommendations or accept subscriptions. This site publishes what the machine learning trading system does and how it was researched; that is the entire scope.
Where can I ask something technical?
Discord, linked below. Questions about XGBoost, feature construction, walk-forward analysis, out-of-sample validation and design are welcome. Personal investment recommendations are not provided.
Support
Keeping this online
Hosting, data feeds and compute are the running costs. If the record or the write-ups have been useful, a small contribution helps keep them published. It buys nothing — there is no tier, no access and no acknowledgement attached.
bc1q2et03nku2kkkqaqz20g7ymvw3wl24njm3ez4lv
Contact
Ask about the method
The Discord server is where I answer questions about how the machine learning trading system is built — features, validation, why a given approach was dropped. It exists so the technical side of this project is open to scrutiny rather than taken on trust.
Risk notice
Not investment advice. The content published on this website, including all trade records, live trades and written analysis, is provided for informational and educational purposes only. It does not constitute investment advice, a personal recommendation, an offer, or a solicitation to buy or sell any financial instrument, and it takes no account of the objectives, financial situation or needs of any individual reader.
Backtested and simulated results. The Frontier historical research track is a backtest study, while the FXBlue account is a separately tracked simulated execution record operating under live market conditions. No real capital is represented by the backtest. Backtested and simulated results have inherent limitations and do not fully reflect real-world execution factors including slippage, partial fills, requotes, financing costs, funding and spread variation. Past performance, whether backtested, simulated or real, is not a reliable indicator of future results.
Risk warning. Trading bitcoin and other digital assets carries a high level of risk and is not suitable for every investor. You may lose some or all of your capital. Leverage increases both gains and losses. Before trading you should carefully consider your objectives, level of experience and appetite for risk, and you should not commit funds you cannot afford to lose. If you are in any doubt, seek advice from an independent and appropriately authorised financial adviser.
No liability. Any decision you take after reading this website is your own. The author accepts no responsibility or liability for any loss or damage arising directly or indirectly from the use of, or reliance on, any information published here. You acknowledge that no financial advice is being provided and that you alone are responsible for the trades you place.
Jurisdiction. This website is not directed at, and its content is not intended for, residents of the United States of America, or any person in any jurisdiction where the publication or availability of this content would be contrary to local law or regulation.
Author. Adrian Velai is a pen name. This website is an independent personal project. It is not affiliated with, endorsed by, or acting on behalf of any broker, exchange or financial institution, and receives no commission, rebate or referral fee from any such party.