Gold Hedge

After a bit of trial and error, mostly error, I think I have the trader back on track.

We lost about $250 just on the NY market session opens over the last few weeks.

I have managed to work out the original parameters that were working initially and hope that this survives longer term.

'Starting over', from 13th April 2026, the trading has been slow and steady.
The table doesn't convert well - sorry.


XAUUSD M5 — TRADING REPORT
Period: 2026-04-13 — 2026-04-14
EA: claude-gold-over-m5 v1.11

[ PERFORMANCE SUMMARY ]
────────────────────────────────────────────────────────────────────────────
Start Balance (2026-04-12 Close) : $754.64
Current Balance (2026-04-14) : $844.14
Total Net Profit : +$89.50
P/L % Change : +11.86%
────────────────────────────────────────────────────────────────────────────
Gross Profit : +$89.67
Gross Loss : -$0.17
Profit Factor : 527.47
Win Rate : 96.15% (25 Wins / 1 Loss)
Total Closed Trades : 26
────────────────────────────────────────────────────────────────────────────
[ SESSION EVENTS & ANALYSIS ]

• STRATEGY STABILITY: The period shows extremely high consistency with a near
linear equity curve. The v1.11 update's trailing SL and session-aware
widening are effectively capturing high-probability scalps.

• LOT SIZE SCALING: On April 13 (17:21–19:47), the EA attempted to scale
risk to 0.03 lots based on high win-rate confidence. These orders were
rejected by the broker ("no money"), indicating the account's free margin
is currently the primary bottleneck for accelerated growth.

• VOLATILITY MANAGEMENT: April 14 showed higher profit per trade in the
later sessions, with the final three trades of the period (#24, #25, #26)
yielding the highest returns ($4.52, $6.13, and $7.03 respectively).

[ CLOSED TRADES LIST ]
────────────────────────────────────────────────────────────────────────────
# Ticket Close Time Dir Open Close P/L Balance
────────────────────────────────────────────────────────────────────────────
01 1880821 2026.04.13 09:49 SELL 4723.48 4721.32 +3.07 $757.71
02 1881113 2026.04.13 10:01 BUY 4728.57 4731.52 +4.18 $761.89
03 1881927 2026.04.13 11:43 SELL 4727.40 4725.02 +3.38 $765.27
04 1882476 2026.04.13 12:29 SELL 4717.50 4715.08 +3.43 $768.70
05 1884446 2026.04.13 14:35 SELL 4718.04 4716.25 +2.54 $771.24
06 1884674 2026.04.13 14:41 SELL 4710.57 4706.80 +5.35 $776.59
07 1884832 2026.04.13 15:11 BUY 4715.08 4718.18 +4.41 $781.00
08 1885104 2026.04.13 15:33 BUY 4713.54 4715.49 +2.77 $783.77
09 1885248 2026.04.13 16:11 BUY 4728.62 4730.53 +2.71 $786.48
10 1885500 2026.04.13 16:50 BUY 4738.74 4740.85 +2.99 $789.47
11 1886200 2026.04.13 16:52 SELL 4730.07 4727.31 +3.91 $793.38
12 1890362 2026.04.14 02:02 BUY 4760.82 4762.61 +2.52 $795.90
13 1890453 2026.04.14 03:00 SELL 4756.04 4753.93 +2.98 $798.88
14 1890776 2026.04.14 03:15 BUY 4759.60 4763.28 +5.19 $804.07
15 1891160 2026.04.14 03:41 BUY 4772.91 4775.31 +3.39 $807.46
16 1891318 2026.04.14 04:00 SELL 4767.95 4765.53 +3.41 $810.87
17 1891735 2026.04.14 04:33 SELL 4767.78 4765.83 +2.76 $813.63
18 1891948 2026.04.14 05:45 BUY 4774.10 4774.48 +0.54 $814.17
19 1892889 2026.04.14 07:19 BUY 4772.74 4772.62 -0.17 $814.00
20 1893422 2026.04.14 08:09 SELL 4770.94 4768.85 +2.95 $816.95
21 1893618 2026.04.14 08:27 SELL 4764.21 4762.07 +3.03 $819.98
22 1893724 2026.04.14 09:09 BUY 4767.31 4769.17 +2.62 $822.60
23 1894118 2026.04.14 09:53 BUY 4777.29 4780.02 +3.85 $826.45
24 1894656 2026.04.14 10:22 SELL 4785.08 4781.87 +4.52 $830.97
25 1895200 2026.04.14 13:38 SELL 4770.49 4766.13 +6.13 $837.10
26 1896273 2026.04.14 14:24 BUY 4774.25 4779.26 +7.03 $844.13
────────────────────────────────────────────────────────────────────────────
[ FOOTNOTE ]
* Lot Scaling attempts (#1886920 to #1888694) resulted in "deleted [no money]".
* Reported balance reflects closing prices and settled P/L in AUD.
* Broker time is UTC+3. Final trade closed at 14:24:50 broker time.
────────────────────────────────────────────────────────────────────────────

Karpathy's Autoresearch and improving TA to determine the next M5 candle close

https://github.com/karpathy/autoresearch

As the current system seems to be going well, 95%+ win rate, I also have decided to look into developing a separate system that only has a high determination of the market direction. The theory is that I can couple this with the current trading system as a confirmation or if good enough, the pilot to direct the system bias for placing trades. Simples...

Below is a snippet of the output that seems to improve over time.
I am using a local gemma:2b model with ollama as this is all that is needed, no using tokens on Sonnet etc.

--- Cycle 14/50 (Phase 3: Gens 208-223) ---
*** New Best! Accuracy: 57.19% | Signals: 334 | Score: 9.67
*** New Best! Accuracy: 57.59% | Signals: 323 | Score: 10.05
[Ollama] Generating correlation report and recalculating search boundaries...
Shifted macd_slow: (1, 41, 10) -> (1, 61, 15)
Shifted bb_period: (3, 23, 5) -> (4, 24, 5)
Shifted atr_mult: (0.0, 0.1, 0.0) -> (0.0, 0.2, 0.0)
[Ollama] Boundary adjustments locked in for the next cycle.
--- Cycle 15/50 (Phase 3: Gens 224-239) ---
*** New Best! Accuracy: 57.54% | Signals: 358 | Score: 10.26
*** New Best! Accuracy: 59.32% | Signals: 322 | Score: 12.18
[Ollama] Generating correlation report and recalculating search boundaries...
Shifted atr_mult: (0.0, 0.2, 0.0) -> (0.0, 0.1, 0.0)
[Ollama] Boundary adjustments locked in for the next cycle.
--- Cycle 16/50 (Phase 3: Gens 240-255) ---
*** New Best! Accuracy: 59.33% | Signals: 359 | Score: 12.39
*** New Best! Accuracy: 58.81% | Signals: 403 | Score: 12.54
[Ollama] Generating correlation report and recalculating search boundaries...

System Overview: The AI-Guided "Predictive Scientist"
This system is a high-speed research engine designed to predict the direction of the next 5-minute price movement in the Gold market (XAUUSD). Instead of relying on human intuition, it acts like a digital laboratory that conducts hundreds of experiments to find the perfect "mathematical recipe" for the current market conditions.
1. How the "AutoResearch" Engine Works
The system uses three main components to find winning strategies:

1. The Toolbox (Technical Indicators): The engine uses standard mathematical tools like Moving Averages (EMA), Momentum (RSI), and Volatility (Bollinger Bands). These are used to "read" the market’s current state.
2. The Scientist (Hierarchical Search): The engine doesn't just guess randomly. It works in phases. First, it tries to find the general trend. Once it has the trend, it fine-tunes the timing. Finally, it adjusts for volatility (how much the price is jumping).
3. The AI Brain (Ollama Advisor): This is the most unique part. We have integrated a local AI (a Large Language Model) that "sits" inside the system. After every round of experiments, the AI looks at the results. It identifies which settings are failing and which are succeeding. It then dynamically shifts the search boundaries, telling the engine, "Stop looking over there; the data suggests the answer is over here."

2. Understanding the Backtesting Process
Backtesting is the process of testing a strategy on historical data to see how it would have performed. However, this system doesn't measure "money" or "profit," because those can be misleading due to fees and luck.
Instead, it measures Predictive Accuracy.

1. It asks a simple question: "Based on the math, will the next 5-minute candle close higher or lower than the current one?"
2. The system then compares every single "guess" against what actually happened.
3. The result is a Win-Rate percentage. If a strategy has a 58% win rate over 5,000 bars, it suggests a significant statistical edge rather than a random fluke.

3. Why This System Isn't "Overfitting"
In trading technology, Overfitting is the biggest danger. It’s like a student who memorizes the answers to a specific test but fails when the questions are slightly changed. An overfitted model looks perfect on past data but crashes in the real world.
We prevent this through four specific safeguards:

1. The 80/20 Training Split: The system is "blindfolded" during research. We give it 80% of the data to find a strategy, and then we test that strategy on the remaining 20% (data it has never seen before). If the strategy fails on that 20%, it is discarded as a "memory fluke."
2. Signal Density Requirements: The system ignores any strategy that only triggers once in a blue moon. It requires a minimum frequency (e.g., at least 5 signals per day). This ensures the results are based on repeated patterns, not one-off lucky events.
3. Linear Correlation Filtering: Our "AI Advisor" looks at the relationship between settings. If a win-rate only goes up when a setting is at a very specific, weird number, the system flags it as suspicious. We look for "smooth" trends where small changes in settings lead to predictable changes in results.
4. Mathematical Floors: As seen in our recent updates, we force the system to maintain a "wide" search area. By preventing the AI from narrowing the search to one specific number, we ensure the strategy is robust enough to handle the "noise" of a real market.

This isn't a "black box" that prints money. It is a computational research tool that uses local AI to navigate millions of mathematical combinations. It filters out luck, demands statistical proof, and constantly evolves its search area to find where the highest probability of directional movement resides.

Current Status (since starting writing this we have gone from 58-59% 'win rate' at predicting the market direction to almost 61%)

The system will save the 'best parameters' as a json which will be used by gemma:2b (or whatever model you use) to search the spaces and get a better score thus, increase the market direction prediction.

YES, if you find these TA parameters useful - fill your boots!

OUTPUT:

{
"timestamp": "2026-04-14 21:56:17",
"generation": 387,
"best_score": 19.650774,
"metrics": {
"win_rate": 60.96,
"total_signals": 251,
"signals_per_day": 11.4,
"up_win_rate": 55.71,
"down_win_rate": 67.57,
"consec_correct": 9,
"consec_incorrect": 6,
"score": 19.6508
},
"params": {
"ema_fast": 10,
"ema_slow": 13,
"macd_fast": 10,
"macd_slow": 23,
"macd_signal": 2,
"adx_period": 4,
"adx_thresh": 24.7,
"rsi_period": 5,
"rsi_oversold": 29.0,
"rsi_overbought": 79.0,
"stoch_k": 9,
"stoch_d": 3,
"stoch_smooth": 9,
"bb_period": 24,
"bb_stdev": 1.0,
"atr_period": 20,
"atr_mult": 0.006112762124423629
},
"current_search_space": {
"ema_fast": [
1,
21,
5
],
"ema_slow": [
1,
61,
15
],
"macd_fast": [
1,
21,
5
],
"macd_slow": [
1,
61,
15
],
"macd_signal": [
1,
5,
1
],
"adx_period": [
1,
21,
5
],
"adx_thresh": [
4.1,
24.7,
5.0
],
"stoch_k": [
6,
14,
2
],
"stoch_d": [
1,
9,
2
],
"stoch_smooth": [
1,
9,
2
],
"bb_period": [
4,
24,
5
],
"bb_stdev": [
1.0,
5.0,
1.0
],
"atr_period": [
5,
25,
5
],
"atr_mult": [
0.0,
0.006112762124423629,
0.0
]
}
}

Lets see if we survive the next 24 hours.

[ PERFORMANCE SUMMARY ]

Start Balance (2026-04-12 Close) : $754.64
Current Balance (2026-04-14) : $877.91
Total Net Profit : +$123.27
P/L % Change : +16.33%

Gross Profit : +$123.58
Gross Loss : -$0.31
Profit Factor : 398.64
Win Rate : 94.29% (33 Wins / 2 Losses)
Total Closed Trades : 35

Attached is the plain text report.

Temporary html report - this will be deleted when I post an updated one in the future.

https://trade-report-2026-04-13-2026-04-14.tiiny.site/

Directional System Update

I have updated the autoresearch and backtester to now added to forecast the direction 15, and 30 minute candle close.

Since I am using the M5 for my system it seemed logical to try and work on 15 minute candle close. The switch can be for 5 minutes, e.g. --horizon 5 or --horizon 30, if you want 30 minute candle close.

My trading system use the RSI 5, 29 and 79 so I have these hard coded and these will not changed during the discovery and search space.

There is a bit more longer time period testing to complete but the results are promising.

COMMANDS:
1. DEBUG: Run a baseline backtest on the current CSV to verify data loading.
python main.py --mode debug --horizon 15

2. MANUAL: Run a backtest using specific parameters from a saved JSON report.
python main.py --mode manual --params output/best_report.json --horizon 15

3. AUTO (Fresh): Start a new hierarchical search session.
python main.py --mode auto --n 50 --horizon 15

4. AUTO (Resume): Continue a previous search using 'best_report.json' as a seed.
python main.py --mode auto --n 100 --resume --horizon 15


UPDATED SCORE: Win Rate = Overall %

{
"timestamp": "2026-04-15 08:43:55",
"generation": 622,
"best_score": 21.1117,
"metrics": {
"win_rate": 71.68,
"total_signals": 113,
"signals_per_day": 12.4,
"up_win_rate": 72.46,
"down_win_rate": 70.45,
"consec_correct": 20,
"consec_incorrect": 4,
"score": 21.1117
},
"params": {
"ema_fast": 65,
"ema_slow": 81,
"macd_fast": 1,
"macd_slow": 10,
"macd_signal": 14,
"adx_period": 67,
"adx_thresh": 14.543097600545,
"rsi_period": 5,
"rsi_oversold": 29.0,
"rsi_overbought": 79.0,
"stoch_k": 68,
"stoch_d": 74,
"stoch_smooth": 44,
"bb_period": 40,
"bb_stdev": 16.659351361493876,
"atr_period": 24,
"atr_mult": 0.0
},

Suffered the first big loss today of -$6.28 around the Asian/London swap.

With 2 losses in the same trading day I'm hoping it won't get worse.

Waiting to see how the London / NY Markets are - can the system weather the storm?

[ PERFORMANCE SUMMARY ]

Start Balance (2026-04-12 Close) : $754.64
Current Balance (2026-04-15) : $903.33
Total Net Profit : +$148.69
P/L % Change : +19.70%

Gross Profit : +$155.84
Gross Loss : -$7.15
Profit Factor : 21.79
Win Rate : 91.11% (41 Wins / 4 Losses)
Total Closed Trades : 45

Looks like tiny site URL link doesn't work. Sorry.

Screenshot of the report instead - just matches the trading history from the 13th of April 2026.

Lets get back on track next week!

We hit the wall mid week showing that the system wasn't able to handle the market volitilaty.

We have reviewed and hopefully improved the system to adapt to next weeks' turbulant market.

Good news is that the updated system now has the dynamic lots based on the confidence. Last two trades were about $12 and $18 profit.

Here is the repo for the Autoresearch and Backtester - FULLY automated with Ollama and Gemma4:e2b so it doesn't chew up your tokens with a subscription.

https://github.com/BeanBagData/Gold-...testing-System

Hello, I'm very interested in your work. I'd like to ask a question related to it. Could this system be transformed into a structure that optimizes real-time data and develops software based on that input?