Now training v2.5.5 — partial close fix · RSI 30 floor · cleaner entries
  Access closed

The AI trading bot that
learns, not repeats.

Reinforcement learning agent trained to discover profitable strategies — starting with your first $100.

📋
Join the waitlist
Be first in line when we open access. We'll reach out personally before launching.
Already have a code? Enter it below ↓
Invite only Application reviewed manually Limited spots
Why AxiomBot

Not another trading bot.

🧠

Learns From Scratch

RL-native — discovers strategies from first principles through thousands of simulated trades. No copied templates, no backtested indicators. Pure learned intuition.

🔍

Transparent Decisions

Visible reward signals and explainable logic, not a black box. Watch the agent train, see every version's scores, and understand why it trades the way it does.

👤

Built for Solo Traders

Starts small ($100), scales with you. No hedge fund required. Designed for independent traders who want AI working for them, not against them.

AXIOM//ACCESS
v2.5.5
🔑
Have an access code?
Access codes are issued to approved members only. Enter yours below to unlock the platform.
Don't have a code? Access is currently invite-only — see plans to apply when we open.
axiombot — training session
LIVE
$ current versionv2.5.5 — training now · bug fixes
$ v2.0 tests passed3/4 · 75%
$ window 1 Aug–Dec 2024+4.64% Sharpe 7.27 ✓
$ window 2 Dec 2024–Apr 2025−3.29% Sharpe −2.32 ✗
$ window 3 Apr–Aug 2025+9.62% Sharpe 6.01 ✓
$ window 4 Aug–Dec 2025+0.92% Sharpe 1.58 ✓
$ next step30-day paper trading
Pricing

Join the waitlist.

🔒 Access is currently closed while the bot trains. Join the waitlist for your plan — we'll send your access code personally when we open.
Free Beta
$0 /mo
Get started with paper trading. Our bot trades on your behalf — no setup, no coding.
  • Up to $500 managed by AxiomBot
  • Paper trading during beta period
  • Live performance dashboard
  • Real-time trade notifications
🔒 ACCESS CLOSED — WAITLIST OPEN
Founder Club
Custom
No deposit cap. For serious traders who want maximum exposure through our bot.
  • Unlimited capital deployed by AxiomBot
  • Dedicated account manager
  • Custom risk parameters
  • Direct line to our team
  • First access to new features
🔒 ACCESS CLOSED — WAITLIST OPEN

All plans include: historical data access, encrypted connections, and full transparency of bot decisions.

FAQ

Questions? Answered.

How is AxiomBot different from other trading bots? +
Most bots use pre-programmed rules or backtested indicators. AxiomBot uses reinforcement learning — it discovers strategies through trial and error across thousands of simulated trades, like how AlphaGo learned to play Go. No templates, no copying.
Is the beta really free? +
Access is currently closed. We are not accepting new members during active training. We are actively training v2.5.5 with critical bug fixes applied. When walk-forward consistently passes we will open a limited number of beta spots — by invitation only.
How much money do I need to start? +
$100 minimum for live trading (Pro plan). During the free beta, you train on simulated capital — no real money at risk. The bot is designed to work with small accounts first and scale from there.
Can I lose money? +
Yes — all trading carries risk. AxiomBot includes a 200-day moving average trend filter to avoid trading in downtrends, but no strategy is risk-free. During beta, you use paper trading only (no real money). We believe in transparency: you can see exactly how the bot performs before putting capital at risk.
What markets does it trade? +
Currently training on US equities (AAPL). We plan to expand to other tickers and eventually crypto and forex as the model matures.
Can I see how the bot is performing right now? +
Yes — the Training Dashboard is live and fully public. Anyone can watch the bot train and see every version's results. What's closed is access to actually use the bot. The dashboard is open so you can judge for yourself whether it's worth waiting for.
ACCESS CURRENTLY CLOSED

Not accepting new members.

We're in active training and validation. Access opens by invitation only — when the bot proves itself, we'll expand carefully.

🔒 Applications are not open at this time
Training Dashboard
Every version, every result — documented live as we build toward a fully autonomous bot.
v2.5.5 TRAINING NOW · BUG FIXES · BEST VERSION YET
Current Version
v2.5.5
Training now · bug fixes
Best Score Ever
-0.34
v2.4.0 · 2.6M timesteps
Best Walk-Forward
2 / 10
v2.5.3 · W2 +20.66% W10 +13.79%
Stocks Training On
5
AAPL MSFT GOOGL NVDA AMZN
Bot Improvement — Every Version
Each bar shows the return % from walk-forward testing. Green = profitable window. The goal is all bars green.
Profitable
Losing
No trades
v2.5.3 — Best result ever · 2/10 ★
v2.5.5 training now · latest results: v2.5.3 · 2/10 passed
What the bot is working on right now
NOW v2.5.5 training · partial close bug fixed (0.10→0.025) · MA50 gate removed · RSI floor 35→30 · cleaner entries
Mar 8 v2.5.3 BEST RESULT · 2/10 passed · W2 +20.66% · W10 +13.79% · loss aversion 1.1→1.3 · Sharpe bonus tuned
Mar 8 v2.4.2 infrastructure fixed · min_trades=50 bug resolved · sequential multi-stock · n_eval_episodes 3→10 · 1/10 passed
Mar 8 v2.4.1 walk-forward · 0/55 failed · discovered min_trades=50 scoring bug was killing all results — not a bot problem
Mar 8 v2.4.0 walk-forward · 0/10 passed · 7/10 profitable · ADX too strict · W8 zero trades · circuit breaker working · avg DD 3.4%
Mar 8 PROP FIRM APPROACH ADOPTED · max 3 changes per version · 2% risk per trade · validate all 5 symbols · stress test crashes
Mar 7 v2.0 PASSED walk-forward · 3/4 profitable · Avg return +2.97% · Window 2 loss cut from −9.97% → −3.29%
Version History
🏆 Major Milestone
Version 1.0 → Version 2.0 — How far we've come
v1.0 — Where we started v2.0 — Where we are now
Trades per test window 0 – 1 1 – 4 ✓
Best training score −0.21 −0.02 ✓
Tests passed 4 / 10 · 40% 3 / 4 · 75% ✓
Avg return (passing windows) mixed / unreliable +5.06% avg ✓
Worst window loss −13.0% −3.29% ✓
Avoids falling markets No ✗ Yes — 200-day filter ✓
Bot behaviour Refused to trade at all Actively managing positions
v2.5.5
Bug fixes — partial close corrected, MA50 gate removed, RSI floor lowered
Mar 8, 2026 · AAPL+MSFT+GOOGL+NVDA+AMZN · Training now
TRAINING
What changed from v2.5.4: Three critical bug fixes. The partial close threshold had crept up to 0.10 (10%) — meaning the bot never triggered partial close because prices rarely move 10% in a single trade. Fixed back to 0.025 (2.5%). The MA50 entry gate was blocking too many valid setups and was removed. RSI floor lowered from 35 to 30 to allow entries in slightly more oversold conditions. These are clean precision fixes targeting the exact reasons entries were being blocked.
🔧 Partial close threshold 0.10 → 0.025 — was never triggering 🔧 MA50 entry gate removed — was blocking too many valid setups 🔧 RSI floor 35 → 30 — allow more oversold entries ✓ All v2.5.3 rules preserved ⚠ Walk-forward results pending
v2.5.3
Best result yet — 2/10 passed, W2 +20.66%, W10 +13.79%
Mar 8, 2026 · AAPL · 2/10 windows passed
2/10
What changed from v2.5.2: Reward shaping tuned to punish losses harder and reward consistency. Loss aversion increased from 1.1 to 1.3 — the agent gets penalized 30% more for every losing day instead of 10%. Sharpe bonus tuned to 0.006 to better reward consistent returns over lucky big wins. Stop loss penalty also increased. Result: best walk-forward ever recorded — W2 returned +20.66% and W10 returned +13.79%.
BEST WINDOWS EVER RECORDED
WINDOW 2
+20.66%
WINDOW 10
+13.79%
✓ Loss aversion 1.1 → 1.3 — punishes losses harder ✓ Sharpe bonus tuned to 0.006 — rewards consistency ✓ Stop loss penalty increased to 0.006 ✓ Best single window ever: +20.66% ⚠ 8/10 windows still failing — more work needed
v2.4.2
Infrastructure fixed — sequential training, eval bug resolved
Mar 8, 2026 · AAPL · 1/10 windows passed
TESTED
What changed from v2.4.1: Critical infrastructure fixes. Discovered that v2.4.1's walk-forward (0/55) wasn't a bot problem — the min_trades threshold was set to 50, meaning a window needed 50 trades to even be scored. Most windows had 5-20 trades. Fixed to min_trades=5. Switched to sequential multi-stock training (one stock at a time) instead of parallel — more stable learning. Increased eval episodes from 3 to 10 for more reliable model saving. Result: 1/10 windows passed for the first time.
🔧 min_trades bug fixed: 50 → 5 — was blocking all scoring 🔧 Sequential multi-stock training — more stable than parallel 🔧 n_eval_episodes 3 → 10 — more reliable best model saves ✓ Walk-forward defaults set: 600/80/80 windows ⚠ Still 3 zero-trade windows — entry filters too strict
v2.4.1
Prop firm grade — smarter entries, proper risk sizing, 5-stock training
Mar 8, 2026 · AAPL+MSFT+GOOGL+NVDA+AMZN · Scoring bug found
0/55
What changed from v2.4.0: ADX loosened from 0.25 to 0.20, RSI widened from 40-72 to 35-75, and proper 2% risk position sizing added. Walk-forward returned 0/55 — but investigation revealed this was a scoring bug, not a bot failure. The min_trades criteria was set to 50, so every window was automatically disqualified. Fixed in v2.4.2.
✓ ADX filter loosened 0.25 → 0.20 ✓ RSI range widened 40-72 → 35-75 ✓ 2% risk position sizing — prop firm standard ✗ Walk-forward 0/55 — caused by min_trades=50 scoring bug, not bot 🔧 Bug fixed in v2.4.2
v2.4.0
Multi-stock training — ADX, RSI, dual MA filters, circuit breaker
Mar 8, 2026 · AAPL+MSFT+GOOGL+NVDA+AMZN · Walk-forward complete
TESTED
What changed from v2.2: The biggest upgrade in AxiomBot's history. Switched from training on one stock (AAPL) to five simultaneously — the bot now learns general market behavior instead of memorizing Apple patterns. Added three new entry filters: ADX (trend strength must be confirmed), RSI (not overbought or oversold), and dual moving average (price above both 50-day and 200-day). Added a circuit breaker that stops trading entirely if down 6% from peak. Added cooldown after 3 consecutive stop losses. Best training score ever: -0.34 at 2.6M timesteps.
WALK-FORWARD RESULTS — 10 WINDOWS
W1
-0.67%
W2
-5.30%
W3
+4.92%
W4
-1.26%
W5
+1.82%
W6
+6.53%
W7
+0.27%
W8
0 trades
W9
+0.02%
W10
+8.57%
7/10 windows profitable · Avg return +1.49% · Avg DD 3.4% · Issue: ADX too strict → v2.4.1 fix
✓ Multi-stock training: AAPL, MSFT, GOOGL, NVDA, AMZN ✓ ADX trend strength filter ✓ RSI overbought/oversold filter ✓ Dual MA filter: price above 50-day AND 200-day ✓ Circuit breaker: stops trading if down 6% from peak ✓ Cooldown: sit out 5 bars after 3 consecutive stop losses ⚠ ADX 0.25 too strict — W8 had zero trades — fixed in v2.4.1
v2.2
Partial close — once up +2%, the trade cannot lose money
Mar 7, 2026 · AAPL · Walk-forward complete
TESTED
What changed from v2.1: A single new rule that makes every winning trade safer. The moment a trade reaches +2% profit, the bot automatically sells half the position to lock in guaranteed profit, then moves the stop loss to the entry price. From that point it is mathematically impossible to lose money on the trade — the worst outcome is breaking even on the second half while keeping the first half's profit.
✓ Partial close at +2% — 50% of position locked in ✓ Stop loss moves to entry after partial close ✓ Impossible to lose on any trade that hits +2% ✓ Remaining 50% still runs to +4% for max profit ✓ All v2.1 rules preserved ⚠ Walk-forward 0/10 — win rate too low — led to v2.4.0 overhaul
v2.1
Hard exits — bot no longer has a choice on stop loss or take profit
Mar 7, 2026 · AAPL · Built, training as base for v2.2
TRAINING
What changed from v2.0: Four loopholes closed. In v2.0 the bot could ignore a −2% loss and keep holding. It could ride a +6% winner all the way back to zero. It could hold a position while the entire trend turned against it. All of that is now physically forced — no override.
🚪 Forced MA Exit
v2.0 stopped new buys in downtrends. v2.1 also kicks the bot out if it's already holding when trend turns down.
🛑 Hard Stop Loss
Before, −2% was a soft penalty the bot could ignore. Now it's a forced close. Worst a single trade can do is roughly −2%.
💰 Hard Take Profit
At +4% the bot is forced to close. Before it could ride a +6% winner all the way back to 0%.
📈 Exponential Inaction
Sitting in cash for 20 bars hurts exponentially more than 10 bars — creates urgency to find and enter trades.
✓ Forced MA exit — sells if trend turns while holding ✓ Hard stop loss at −2% — no override ✓ Hard take profit at +4% — locks in gains ✓ Exponential inaction penalty ✓ Rolling Sharpe bonus — rewards consistency ✓ Fresh model — old weights deleted ⚠ Walk-forward results pending
v2.0
Trend Awareness — 200-day filter · Window 2 loss cut by 67%
Mar 7, 2026 · AAPL · 2.5 avg trades/window
PASSED
What happened: The 200-day trend filter worked. Window 2 (Dec 2024 → Apr 2025) — the window that lost money in every single previous version — dropped from −9.97% all the way to −3.29%. The filter stopped the bot buying into the Apple downturn. Three windows profitable, Sharpe ratios strong in the winning windows (7.27 and 6.01). Window 2 is still the weak spot but is now manageable rather than catastrophic.
3 / 4
Tests Passed
+2.97%
Avg Return
3.64
Avg Sharpe
4.87%
Avg Drawdown
2.5
Avg Trades
AUG–DEC 2024
+4.64%
Sharpe 7.27
DD 1.41% · 4 trades
DEC 2024–APR 2025
−3.29%
Sharpe −2.32
DD 9.69% · 2 trades
APR–AUG 2025
+9.62%
Sharpe 6.01
DD 5.61% · 1 trade
AUG–DEC 2025
+0.92%
Sharpe 1.58
DD 2.75% · 3 trades
Aug–Dec 2024 · +4.64%
Dec 2024–Apr 2025 · −3.29%
Apr–Aug 2025 · +9.62%
Aug–Dec 2025 · +0.92%
Aug–Dec 2024Dec 2024–Apr 2025Apr–Aug 2025Aug–Dec 2025
✓ 200-day trend filter added ✓ Window 2 loss cut 67%: −9.97% → −3.29% ✓ Sharpe 7.27 in Window 1 — strongest ever ✓ Avg drawdown under 5% ✓ All v1.9 penalties preserved ✗ Window 2 still losing — further work needed
v1.9
Bot trades actively for the first time
Mar 7, 2026 · AAPL · 3.25 avg trades/window
PASSED
What happened: We cranked up every penalty — holding a position too long, sitting in cash, letting a loss run past 2%. It worked. Trades jumped from 1–2 per window to 3.25 on average. Training episodes shortened to 56 bars — the bot was entering, exiting and re-entering trades regularly for the first time ever. Also hit the best score in project history: −0.02, three separate times.
3 / 4
Tests Passed
3.25
Avg Trades
−0.02
Best Score
4.99%
Avg Max Loss
Aug–Dec 2024 · +4.22%
Dec 2024–Apr 2025 · −9.97%
Apr–Aug 2025 · +5.76%
Aug–Dec 2025 · +1.33%
Aug–Dec 2024Dec 2024–Apr 2025Apr–Aug 2025Aug–Dec 2025
✓ Position hold threshold: 15 → 8 bars ✓ Cash inaction threshold: 10 → 5 bars ✓ Bigger reward for closing a winning trade ✓ Harder penalty for letting losses run ✗ Dec 2024–Apr 2025 window still losing every run
v1.8
First real trade management — exits and re-enters
Mar 7, 2026 · AAPL · 2 avg trades/window
PARTIAL
What happened: Every version before this made exactly 1 trade per window — buy once, hold forever. We added a new rule: the longer you hold the same position without closing it, the more points you lose. Trades doubled from 1 to 2 per window. Small step, but the first time the bot showed real active trade management.
3 / 4
Tests Passed
2
Avg Trades
−0.06
Best Score
Aug–Dec 2024 · +1.80%
Dec 2024–Apr 2025 · −8.77%
Apr–Aug 2025 · +8.99%
Aug–Dec 2025 · +0.35%
Aug–Dec 2024Dec 2024–Apr 2025Apr–Aug 2025Aug–Dec 2025
✓ Penalty added for holding same position too long ✓ Bonus reward for closing a winning trade ✓ Extra penalty for letting a loss exceed −2% ✗ Only 2 trades — needed much more active trading ✗ Dec 2024 window still losing
v1.7
First walk-forward pass — still only 1 trade per window
Mar 6, 2026 · Fresh start · Daily prices · 4 years of AAPL
PASSED
What happened: Deleted the old model and started completely fresh. Switched from hourly price snapshots to one clean daily summary — less noise, more signal. Trained on 4 full years of Apple stock covering bull runs, crashes, and volatile periods. Passed walk-forward testing for the first time ever. However the bot was only making 1 trade per window — buying once to escape the cash penalty then holding forever.
3 / 4
Tests Passed
1
Avg Trades
−0.08
Best Score
✓ Daily price bars — cleaner data ✓ 4 years of history (1,825 days) ✓ First ever walk-forward pass ✗ Only 1 trade per window — no real management ✗ Dec 2024 window consistently losing
v1.5
Bot finally started making different decisions
Mar 6, 2026 · All fixes combined · AAPL
NO EDGE
What happened: For the first time the bot started making genuinely different decisions in different market situations. We added random starting points during training so it practices in all types of markets — not just one stretch of history — and a bonus for trying new strategies. Real progress, but it couldn't push into positive scores. The learning rate was set too high, causing it to overshoot every time it got close.
✓ Random starting points in training ✓ Exploration bonus added ✓ Stronger penalty for doing nothing ✗ Couldn't break into positive scores ✗ Learning rate too high — kept overshooting
v1.3
Great in rising markets — crashes in downturns
Mar 6, 2026 · 10 time windows · MSFT
NO EDGE
What happened: Made the laziness penalty 4× stronger and it forced the bot to actually make trades. In calm rising markets it was brilliant — one window returned +17.25%. But when the market got volatile and dropped in late 2025, the bot had no idea what to do. It had only ever practiced in rising markets so a downturn was completely foreign to it.
5 / 10
Tests Passed
+17.25%
Best Window
−16.9%
Worst Window
✓ Laziness penalty 4× stronger ✓ No empty test windows ✗ Crashed hard in falling markets ✗ Never trained on bear market conditions
v1.1
Discovered why the bot refused to trade
Mar 6, 2026 · First real-world test · 10 windows · AAPL
NO EDGE
What happened: First proper real-world test — 10 different time periods of Apple stock. Failed badly, losing money in 6 out of 10 with zero trades in 3 windows. But this failure was valuable. It revealed the core problem: the bot had learned that doing absolutely nothing was the safest choice, because there was no punishment for inaction. Like a student who leaves every exam answer blank.
4 / 10
Tests Passed
3
Zero-Trade Windows
✓ First real-world test completed ✗ 3 windows with zero trades at all ✗ No punishment for doing nothing → Root cause found — led to laziness penalty in v1.2
v1.0
The very first bot — looks great in practice, fails in reality
Mar 6, 2026 · Built from scratch · Apple stock
NOT READY
What happened: The very first version. In its own practice environment it looked incredible — turning simulated $100,000 into $170,000. But when tested on real historical data it had never seen before, it fell apart. The bot had memorised specific moments in market history rather than actually learning how to trade. Like a student who memorises the answer sheet instead of understanding the subject.
+70%
Practice Return
4 / 10
Real Tests Passed
3.98
Best Test Score
Window 1
Window 2
Window 3 — 0 trades
Window 4
Window 5
Window 6 — 0 trades
Window 7 — 0 trades
Window 8
Window 9
Window 10
Window 1Window 5Window 10
✓ AI brain built from scratch (PPO) ✓ Stock market simulator built ✓ Losses punished more than gains rewarded ✗ Memorised history — didn't truly learn to trade
🔒
Dashboard is private
Enter your access code to unlock your personal dashboard.
🔒
Admin access required