Experimental MT5 EA using Gemini Flash API - GeminiCommanderEA_v1

Update coming soon - after I finish testing a few new ideas to improve the script.

What is already included (I am up to Version 103) a circuit breaker when the draw down becomes too significant it stops the training and then this data about the trades leading up to the circuit breaker being tripped is also sent to Gemini for analysis as well.
The AI's analysis is currently ephemeral; it learns from cycle to cycle but starts fresh with every new script version. By creating a persistent log of key findings, we can give the AI a "long-term memory," allowing it to identify what has historically worked best across different architectures and feature sets.



Key Enhancements:

1. Persistent Memory File (framework_memory.jsonl): A new file will be created in your Results folder. After each complete run, the script will automatically append a new entry summarizing the version, the final performance, the configuration of the best cycle, and the top-performing features.
2. New "Pre-Flight Check" Stage: At the very beginning of a new run, the script will now:
   1. Read the historical framework_memory.jsonl.
   2. Send this entire history to Gemini with a new, strategic prompt.
   3. This new prompt asks the AI to act as a lead strategist, analyse the successes and failures of all past versions, and recommend the most promising starting configuration for the current run.

   
   

3. AI-Driven Initialization: The hardcoded current_config is now just a fallback. The script will use the configuration recommended by the AI during the "Pre-Flight Check" as its starting point, allowing each new version to be "smarter" than the last.

This transforms the framework from a static tool into a truly evolving intelligence that builds upon its own history.

Evolution Summary: V90.1 to V130

This is my final version I'll be sharing, it has everything you need to find and develop you own strategy for the MQL5 platform.

Just set it to '280 runs' so it stays within the daily free Google Flash 2.0 API limit and within 24 hours you'll have a strategy - probably sooner than that.

If you have questions it is better to message me directly.

The journey from V90.1 to V130 marks the transformation of the framework from a single-strategy backtester into an intelligent, multi-strategy system capable of learning and adapting over time. Key milestones include the integration of SHAP for model interpretability, the introduction of a Daemon Mode for continuous operation, the development of a Strategy Playbook, and the groundbreaking implementation of AI-driven Hybrid Strategy Synthesis and Strategic Health Reviews. The framework's robustness was also significantly enhanced by overhauling the memory system to prevent data corruption and improving the AI's response handling.

Attachments:

1. User manual - short but will explain how it works, I just suggest copying and pasting the script into Google AI Studio and asking it questions if you want to know more.
2. End_To_End_Advanced_ML_Trading_Framework_PRO_V130_Report_Nicknames.txt just rename it and save it as a .py file
3. Pyrope_TrendFollower_20250610-232956_run.txt (this is actually a log) example of the final output when CONTINUOUS_RUN_HOURS = 2

I have tried to ensure that this is a fully developed script that anyone can use out of the box - apart from the .csv files which you can get from MQL5 Desktop Terminal. This requires multi-pair and multi-timeframes for the training so just adding one asset will not be beneficial.
IF there are more bugs, just copy the script into Google AI Studio or the Gemini Web App and paste in the trace error.

One could easily convert this into a desktop app to have a GUI etc. I am running this on a Linux VPN in a terminal - this is designed for Windows.

Use this how you want. Add more strategies, how? copy and paste the ones into something like Perplexity and ask it to expand the strategies -simples.
Yes, you can integrate this into real live trading and it will train on news feeds.
Possibilities are endless.

Version-by-Version Changelog

V90.1 - SHAP Fix

1. Fix: Addressed a foundational bug in the calculation and display of SHAP (SHapley Additive exPlanations) values, ensuring accurate and reliable feature importance reporting. This version established the baseline for model interpretability.

V112 - Playbook & Daemon Control

1. Feature: Introduced the Strategy Playbook, a major architectural shift allowing the AI to choose from a diverse set of predefined trading archetypes (e.g., TrendFollower, MeanReversion).
2. Feature: Implemented Daemon Mode, enabling the script to run continuously for a set duration or number of iterations, managed by new CONTINUOUS_RUN_HOURS and MAX_RUNS controls.
3. Feature: The AI was empowered to select the master strategy at the start of a run and tune its parameters between cycles, marking the first step towards an adaptive framework.

V113 & V114 - Usability and API Key Fixes

1. Fix (V113): Corrected a NameError related to the playbook and fixed an issue where the .env file was not being loaded correctly.
2. Feature (V114): Added a manual input override for the GEMINI_API_KEY. If the key is not found in the .env file, the script now prompts the user to paste it directly, improving usability.

V115 to V118 - Daemon Mode Refinements & Bug Fixes

1. Feature (V115): Adjusted timing delays to better manage API rate limits. The inter-cycle delay (within a run) was increased to 5 minutes, while the inter-run delay was shortened.
2. Feature (V116): Implemented an interruptible 10-second countdown between daemon runs, allowing the user to gracefully stop the process with Ctrl+C without corrupting the run in progress.
3. Docs (V117): Added detailed comments to the daemon controls to clarify the timing logic and confirm that a run, once started, will always complete.
4. Feature (V118): Established two distinct run modes. The default mode runs once with a 5-second inter-cycle delay, while Daemon Mode uses the longer, more robust 5-minute delay and enables advanced AI features.

V119 to V121 - The Self-Evolving Framework

1. Feature (V119): Implemented the Hybrid Strategy Synthesis blueprint.
   1. The playbook was externalized to strategy_playbook.json.
   2. The AI can now analyse top-performing runs and be prompted to create entirely new hybrid strategies, which are then saved back to the playbook.

   
   

2. Feature (V120): Implemented Phase 1 of the Strategic Review.
   1. The framework now analyses the entire run history to generate a "Health Report" for each strategy, calculating its ChronicFailureRate and CircuitBreakerFrequency.
   2. The AI's pre-flight prompt was enhanced to use this report to make more informed decisions.

   
   

3. Feature (V121): Implemented Stagnation Detection.
   1. The Strategic Review now detects if a strategy has been used for 3+ consecutive runs without performance improvement or with high failure rates.
   2. The AI is explicitly instructed to penalize stagnating strategies to force exploration.

   
   

4. Fix (V121): Resolved a KeyError that occurred when analysing historical runs from older script versions.

V122 to V130 - Robustness and Reporting Overhaul

1. Fix (V122-123): Corrected a missing import random statement and fixed a TypeError related to saving date objects to JSON, which was causing file corruption.
2. Feature (V124): Overhauled the memory system for robustness. Historical runs are now saved to a historical_runs.jsonl file, where each run is a single line. This prevents a crash from corrupting the entire history, as the loader can now safely skip malformed lines.
3. Fix (V125-128): Implemented a recursive "flattening" function to correctly parse AI-suggested parameters, making the script resilient to variations in the AI's JSON output format. Added validation to the hybrid synthesis process to prevent malformed strategies from being saved.
4. Fix (V129): Restored the analyze_cycle_and_suggest_changes method that was accidentally removed in a prior update.
5. Feature (V130): Introduced Report Nicknames.
   1. The framework now uses the AI to assign a unique, one-word codename (e.g., Apollo, Vortex) to each script version.
   2. These nicknames are used in reports for a cleaner, more readable layout.
   3. A nickname_ledger.json file was created to store the version-to-nickname mappings, and reports now include a key for traceability.

   
   

By version V130, the script has evolved into a highly robust, self-adapting system that not only tests strategies but also actively diagnoses their long-term health, forces exploration away from failing patterns, and synthesizes novel hybrid strategies based on its own performance history.

I was going thru few last version of py. I can't find anywhere to export data in CSV. Do you have separate py script? i am curious about trying AI/ML but learning python

Sorry everyone, I just found a couple of things that needed fixing.

Script Update: V1.31
Based on the analysis, I have made the following improvements in version 1.31.

1. V1.31 UPDATE (Robustness & Traceability):
   1. FIX - Folder Naming: Result folders are now named using the format {Nickname}_V{Version} (e.g., Pyrope_V131), ensuring every run is traceable and folders are sorted chronologically by version.
   2. FIX - AI Response Parsing: The script is now more robust at parsing AI configurations. It can correctly handle nested parameters blocks that the AI sometimes returns, preventing errors between daemon runs.
   3. FIX - Hybrid Strategy Prompt: The AI prompt for generating new hybrid strategies has been heavily reinforced to demand parameter ranges in the correct format (e.g., [100, 200]), preventing the creation of invalid strategies.
   4. IMPROVEMENT - Pre-Flight Prompt: The pre-flight check prompt now explicitly asks the AI to provide selected_features, reducing the chance of invalid configurations and reliance on fallbacks.

There is not data that gets exported to csv. It only uses csv files to train from.

I got mine from the MQL5 Desktop Terminal, but the main thing is to ensure that the headers are correct.

The only exports are the files created from the report:

1. Report .txt
2. SHAP Summary .png
3. Run Log .log
4. Equity Curve .png

If you wanted the parameters of the strategies they are in the strategy_playbook.json
The other json files can be converted and likewise the strategy_playbook.json can be converted to c.csv if that's what you're after?

I got the CSV bar data from MT5 Terminal.

1. Goto View and Symbols
2. (this will open in Specifications) but select the tab Bars
3. Once in Bars, choose the currency and timeframe and date range
4. Press the Request and Export

Unfortunately the M15 csv are too big to upload ~6mb.

thanks for with images too. It's in History centre in MT4 and i try to find in MT5 but never knew it's hidden there so google it and it came up it's been removed so made my own script that i will change as i go along attaching here that script to download data in one go instead of doing manually if anybody needs it.
Good to know that option still there for exporting, thanks
[Script updated]

The moment you change one thing...

Script Update: V1.32 + the requirements below

This version introduces a critical new layer of defense: AI Parameter Sanitization. It also incorporates the fixes from the previous version which you hadn't run yet.

1. V1.32 UPDATE (AI Guardrails & Sanitization):
   1. NEW - AI Parameter Sanitizer: The framework now actively sanitizes numerical inputs from the AI before validation.
      1. It checks for invalid MAX_DD_PER_CYCLE values (like 70) and automatically converts them to the correct float format (e.g., 0.70), preventing crashes.
      2. It clamps LOOKAHEAD_CANDLES and OPTUNA_TRIALS within a reasonable min/max range to stop the AI from suggesting absurd values.

      
      
      
      

   2. FIX - Folder Naming (from V1.31): Result folders are now named using the format {Nickname}_V{Version} for better traceability.
   3. FIX - AI Response Parsing (from V1.31): The script now correctly handles the nested parameters block in AI responses, improving stability between daemon runs.
   4. FIX - Hybrid Strategy Prompt (from V1.31): The AI prompt for creating hybrids is now much stricter about the required format for parameter ranges, and the script will discard any invalid hybrids the AI creates.

Requirements
matplotlib==3.8.4
numpy==2.3.0
optuna==4.3.0
pandas==2.3.0
pydantic==2.11.5
python-dotenv==1.1.0
Requests==2.32.4
scikit_learn==1.7.0
shap==0.47.2
xgboost==3.0.2

User Requested FIX

Problem Identified:
The start date for the testing (cycles) was hardcoded to 2024-01-01 and users would have a problem if this date was not part of their own CSV set.


The number of cycles is calculated based on three parameters set in the fallback_config dictionary inside the main() function:

1. FORWARD_TEST_START_DATE: This is hardcoded to "2024-01-01".
2. RETRAINING_FREQUENCY: This is hardcoded to '90D', meaning a new cycle starts every 90 days.
3. End Date of Data: Your data files end around June 2025.

The script divides the total time from January 2024 to June 2025 into 90-day chunks, which always results in exactly 6 cycles. The AI influences the parameters within each cycle (like drawdown or features) but not the number of cycles itself.

How It Works Now (V1.33)

1. The script first finds the earliest timestamp available across all your loaded CSV files.
2. It then adds the duration of the initial TRAINING_WINDOW (e.g., '365D') to that earliest date.
3. The walk-forward testing period now automatically begins right after this initial training data, ensuring the framework always uses the maximum amount of available history for training before the first cycle.

With this update, the number of cycles will now change dynamically based on the date range of the CSV files you use.

Script Update: V1.33

1. V1.33 UPDATE (Dynamic Start Date):
   1. CRITICAL FIX - Dynamic Test Start Date: Removed the hardcoded FORWARD_TEST_START_DATE. The framework now automatically calculates the optimal start date for walk-forward testing based on the earliest date in the provided CSV data and the length of the TRAINING_WINDOW.

Another User Request Fix

Problem: I don't want to use the timeframes you've hardcoded into the script and I don't want to enter in each of my csv files one at a time.

NOTE: The csv file naming convention rules apply.

I've updated the script to be fully dynamic in its handling of timeframes. It no longer assumes you are using M15, H1, and D1.
How It Works Now (V1.34)

The framework now intelligently discovers and assigns roles to the timeframes based on your CSV filenames.

1. Automatic Detection: The script reads the filenames in your data directory (e.g., EURUSD_**H4**_...csv, EURUSD_**M5**_...csv). It automatically creates a list of all unique timeframes it finds ([H4, M5, D1]).
2. Role Assignment: It then sorts these timeframes from lowest to highest.
   1. The lowest timeframe is assigned the base role (the one where trades are simulated).
   2. The second lowest is assigned the medium context role.
   3. The third lowest is assigned the high context role.

   
   

3. Dynamic Feature Engineering: The feature engineering process now uses these dynamic roles. Instead of looking for a hardcoded D1 column, it looks for the column associated with the high role, whatever that may be.

This makes the entire system flexible. You can now use any combination of timeframes like M5/H1/D1 or M1/M15/H4, and the script will adapt without requiring any code changes.

Script Update: V1.34

1. V1.34 UPDATE (Dynamic Timeframe Handling):
   1. NEW - Timeframe Detection: The framework no longer hardcodes M15, H1, and D1. It now dynamically detects all timeframes present in the data folder by reading the CSV filenames.
   2. NEW - Dynamic Role Assignment: The script automatically sorts the detected timeframes and assigns them roles: base (lowest), medium, and high (highest), which are then used for feature engineering.
   3. IMPROVEMENT - Fully Adaptive: This change makes the feature engineering and context analysis fully adaptive to any set of timeframes (e.g., M1, M5, M30, H4), as long as the file naming convention (PAIR_**TF**_...) is followed.

Script Update: V1.35

1. V1.35 UPDATE (Default Playbook Fix):
   1. CRITICAL FIX - AttributeError: Corrected the initialize_playbook function by defining the feature-set lists (TREND_FEATURES, SESSION_FEATURES, etc.) locally. This resolves the crash that occurred when the script attempted to create a new, default strategy_playbook.json file.
   2. IMPROVEMENT - Restored Default Strategies: The default playbook now includes the full set of original strategies (TrendFollower, MeanReversion, VolatilityBreakout, etc.) for a more comprehensive starting point.

One last note (I'll update the script later).

Summary of Flaws Based On Some Recent Results:

Unrealistic Compounding: The position sizing model doesn't account for real-world market limitations and scales to absurd levels.

Perverse AI Learning: The AI Strategist creates a positive feedback loop of risk, increasing drawdown limits after failures instead of reducing risk.

High Volatility Masked by Flawed Gains: The strategy is extremely risky, with a low 35% win rate and a massive 54% maximum drawdown. These terrible risk metrics are overshadowed by the fictional, exponentially compounded profits.

Over-Optimization: The constant re-tuning every 90 days with a high number of Optuna trials, combined with the AI's parameter changes, leads to a system that is perfectly curve-fit to the recent past but is clearly unstable, as evidenced by the frequent circuit breaker trips

- I don't know but still messes up all path. Even data at root folder. (csv small data files attached)
- Best way to have data in subfolder - like market_data
- i have update ML_data_collecter script that align with your framework with minor changes

After Looking at code, The way it works and flow as below

Training Data Flow:
# Data Loading & Processing
def load_and_parse_data(self, filenames: List[str]):
# Loads CSV files (OHLCV data) from specified directory
# Handles multiple timeframes (Daily, H1, M15)
# Returns processed DataFrames for each timeframe

# Feature Engineering
def create_feature_stack(self, data_by_tf: Dict[str, pd.DataFrame]):
# Calculates technical indicators (ADX, RSI, etc.)
# Creates multi-timeframe features
# Returns final feature dataset

Model Training:
class ModelTrainer:
def train(self, df_train: pd.DataFrame, feature_list: List[str]):
# Uses XGBoost for classification
# Optimizes hyperparameters with Optuna
# Returns trained model and best threshold

Gemini AI Integration: The script uses Gemini AI for strategy optimization, not for training. It sends requests in these cases:

class GeminiAnalyzer:
def _call_gemini(self, prompt: str):
# Makes API calls to Gemini
# Used for:
# 1. Strategy selection
# 2. Parameter tuning
# 3. Performance analysis
# NOT used for model training

Training Storage:
def save_run_to_memory(config: ConfigModel, new_run_summary: Dict):
# Saves training results to:
# 1. champion.json - Best performing model
# 2. historical_runs.jsonl - All training runs
# 3. strategy_playbook.json - Strategy definitions

So how can we make it to real world trading.
- we have data in csv so are we training on that data and saving model ?
- We sending to gemini api live at end of say every 5-min candle for reinforment learning that with that indicator and time series data

Make as many changes and modifications to the script as you want and I am happy to see it being developed.
It may or may not suit everyone as I am only focusing on 4 currency pairs, other assets etc will require different strategies and ways to approach the analysis for your goals.

The quickest way to update the script to suit anyone is to paste my script into Google AI Studio (unlimited API calls) or
https://gemini.google.com/ because both have the 1,000,000 token limit and you just tell it change, adapt and modify to your own needs with examples etc.

The most I can get out of these two platforms in one shot is about 2,600 lines of code so for me it makes sense to use these for development.

I'll post my own updated version later - mainly includes missing search optimisations forgotten is other versions and API call optimisations (sending more in each call).

What to expect in the next update, including the fixes mentioned previously.

I have added in GNN. There is a pytorch check at the start, see example output.

Since now I created nicknames for the strategies because the full names once hybrids were developed didn't fit well in the reports.
This however created a new API call at start up since there was an pre-analysis conducted which the script would then associate a strategy to try for backtesting.
Given the rate limits, the nickname API call generation was very close to the second that suggested the parameters to use for strategy in the training.
So what I have done is combined the request for the nickname as well as having the script send its own analysis about the M15, H1 and D1 datasets to the one API call which then the response from Flash either confirms or provides its own ideas to apply a strategy.

I have changed the delays between each cycle since the previous changes of not hardcoding the start dates for the cycles and being dynamic to suit anyone's dataset has increased (for me anyway as I use 4 years for each pair) from between 6 and 9 to 13 cycles per run. These (delays) now are less because the Optuna Optimization are adjusted dynamically that can take between 1 and 3 minutes each depending on the number of trials.

I talk about the delays in between the API calls a lot because I want to stay within the Flash 2.0 free trier daily limit and basing everything on 5 minute API calls.

What one could do, is include a timer for to count the total time from an API to when it needs to make another call and add in a dynamic delay to ensure the the 5 minutes in total has expired.

Example below using the output:
From the line - 2025-06-11 22:30:04,246 - INFO - -> Performing Initial AI Analysis & Setup (Single API Call)...
You can record the date and time and when the next call needs to be made.
2025-06-11 22:32:10,327 - INFO - - [SUCCESS] Model training complete
2025-06-11 22:32:10,781 - WARNING - - CIRCUIT BREAKER TRIPPED! Drawdown exceeded 25% for this cycle. Closing all positions.
Then the script would calculate this time difference which is 2 minutes, 6.081 seconds
So the dynamic delay is 5 minutes minus 2 minutes, 6.081 seconds = 2 minutes, 53.919 seconds
Then you have the delay before the next call
Delay of 2 minutes, 53.919 seconds before the next API call
Then the call is made and returned below
2025-06-11 22:32:13,549 - INFO - - AI suggests updating params: {'description': "The previous cycle hit the maximum drawdown, triggering the 'Circuit Breaker'. To mitigate risk and improve stability, I am reducing the maximum drawdown allowed per cycle, shortening the lookahead window, reducing the number of Optuna trials, and simplifying the model by selecting fewer features.", 'MAX_DD_PER_CYCLE': 0.15, 'LOOKAHEAD_CANDLES': 40, 'OPTUNA_TRIALS': 50, 'selected_features': ['VOL', 'ATR', 'RSI', 'bollinger_bandwidth', 'stoch_k', 'momentum_10', 'candle_body_size', 'is_doji', 'hour', 'market_volatility_index']}

After writing this I just plugged this example into Gemini and the new script have this feature.

1. Dynamic 5-Minute API Delay: I've implemented a new APITimer class. This class wraps all calls to the Gemini API. It records the timestamp of each request and, before making the next one, it calculates the time elapsed since the last call. It then dynamically waits for the remainder of the 5-minute (300-second) interval, ensuring a consistent time between the start of each API request.

[EXAMPLE OUTPUT]
2025-06-11 22:30:00,250 - INFO - PyTorch and PyG loaded successfully. GNN module is available.
2025-06-11 22:30:00,300 - WARNING - 'strategy_playbook.json' not found. Seeding a new one with default strategies at: C:\Users\Sam\Documents\Development\Results\strategy_playbook.json
2025-06-11 22:30:00,300 - INFO -
============================== STARTING SINGLE RUN ==============================
2025-06-11 22:30:00,300 - INFO - -> Loading Nickname Ledger...
2025-06-11 22:30:00,300 - INFO - Successfully loaded GEMINI_API_KEY from environment.
2025-06-11 22:30:00,300 - INFO - -> Stage 1: Loading and Preparing Multi-Timeframe Data...
2025-06-11 22:30:01,320 - INFO - - Processed Daily: 5,612 rows for 4 symbols.
2025-06-11 22:30:01,385 - INFO - - Processed H1: 134,656 rows for 4 symbols.
2025-06-11 22:30:01,575 - INFO - - Processed M15: 400,692 rows for 4 symbols.
2025-06-11 22:30:01,575 - INFO - [SUCCESS] Data loading complete. Detected timeframes: ['Daily', 'H1', 'M15']
2025-06-11 22:30:01,590 - INFO - Dynamically determined timeframe roles: {'base': 'M15', 'medium': 'H1', 'high': 'Daily'}
2025-06-11 22:30:01,590 - INFO - -> Stage 2: Engineering Features from Multi-Timeframe Data...
2025-06-11 22:30:02,505 - INFO - - Calculating HTF features for H1...
2025-06-11 22:30:02,710 - INFO - - Calculating HTF features for DAILY...
2025-06-11 22:30:03,540 - INFO - - Merged data and created features. Final dataset shape: (400240, 39)
2025-06-11 22:30:03,540 - INFO - [SUCCESS] Feature engineering complete.
2025-06-11 22:30:04,246 - INFO - --- STRATEGIC REVIEW: Analyzing long-term strategy health...
2025-06-11 22:30:04,246 - INFO - --- STRATEGIC REVIEW: Insufficient history for a full review.
2025-06-11 22:30:04,246 - INFO - -> Performing Initial AI Analysis & Setup (Single API Call)...
2025-06-11 22:30:08,245 - INFO - - Initial AI Analysis and Setup complete.
2025-06-11 22:30:08,250 - INFO - - Saved new nickname 'Prometheus' to ledger.
2025-06-11 22:30:08,250 - INFO - --- Run Initialized: Prometheus | Strategy: TrendPullback ---
2025-06-11 22:30:08,250 - INFO - --- AI Analysis Notes: The market exhibits moderate trending behavior across the analyzed assets. Correlations are relatively high, particularly between EURUSD and GBPUSD. ADX values suggest the presence of trends, but not excessively strong ones. Overall, a mixed environment favoring trend-following strategies that can adapt to pullbacks or short-term range-bound conditions. ---
2025-06-11 22:30:08,250 - INFO - Walk-forward analysis will run for 13 cycles.
2025-06-11 22:30:08,255 - INFO -
--- Starting Cycle [1/13] ---
2025-06-11 22:30:08,265 - INFO - - Dates | Train: 2021-05-18-2022-05-18 | Test: 2022-05-19-2022-08-17
2025-06-11 22:30:08,265 - INFO - - Generating trade labels with Regime-Adjusted Barriers...
2025-06-11 22:30:12,826 - INFO - - Starting model training using strategy: 'Enters on pullbacks (RSI) in the direction of an established long-term trend.'
2025-06-11 22:30:12,851 - INFO - - Starting hyperparameter optimization with risk-adjusted objective (50 trials)...
> Optuna Optimization: Trial 47/50 | Best Score: 0.1599
2025-06-11 22:31:55,166 - INFO - - Optimization complete. Best Objective Score: 0.1599
2025-06-11 22:31:55,166 - INFO - - Best params: {'n_estimators': 600, 'max_depth': 5, 'learning_rate': 0.0525418710147012, 'subsample': 0.777979844263547, 'colsample_bytree': 0.8343384567360479, 'gamma': 2.0170238116663297, 'reg_lambda': 6.991482287697218e-05, 'alpha': 0.0002744212002074299}
2025-06-11 22:31:55,166 - INFO - - Tuning classification threshold for F1 score...
2025-06-11 22:32:01,296 - INFO - - Best threshold found: 0.30 (F1: 0.3936)
2025-06-11 22:32:01,296 - INFO - - Training final model...
2025-06-11 22:32:09,036 - INFO - - Generating SHAP feature importance summary...
2025-06-11 22:32:09,041 - INFO - - Subsampling data for SHAP from 88066 to 2000 rows.
2025-06-11 22:32:10,317 - INFO - - SHAP summary generated successfully.
2025-06-11 22:32:10,327 - INFO - - [SUCCESS] Model training complete.
2025-06-11 22:32:10,781 - WARNING - - CIRCUIT BREAKER TRIPPED! Drawdown exceeded 25% for this cycle. Closing all positions.
2025-06-11 22:32:13,549 - INFO - - AI suggests updating params: {'description': "The previous cycle hit the maximum drawdown, triggering the 'Circuit Breaker'. To mitigate risk and improve stability, I am reducing the maximum drawdown allowed per cycle, shortening the lookahead window, reducing the number of Optuna trials, and simplifying the model by selecting fewer features.", 'MAX_DD_PER_CYCLE': 0.15, 'LOOKAHEAD_CANDLES': 40, 'OPTUNA_TRIALS': 50, 'selected_features': ['VOL', 'ATR', 'RSI', 'bollinger_bandwidth', 'stoch_k', 'momentum_10', 'candle_body_size', 'is_doji', 'hour', 'market_volatility_index']}
2025-06-11 22:32:13,554 - INFO - Pausing for 30 seconds between cycles...
2025-06-11 22:32:43,566 - INFO - --- Cycle complete. PNL: $117,045.80 | Final Equity: $217,045.80 | Time: 155.31s ---
2025-06-11 22:32:43,566 - INFO -
--- Starting Cycle [2/13] ---
2025-06-11 22:32:43,586 - INFO - - Dates | Train: 2021-08-16-2022-08-16 | Test: 2022-08-17-2022-11-15
2025-06-11 22:32:43,586 - INFO - - Generating trade labels with Regime-Adjusted Barriers...
2025-06-11 22:32:47,831 - INFO - - Starting model training using strategy: 'Enters on pullbacks (RSI) in the direction of an established long-term trend.'
2025-06-11 22:32:47,866 - INFO - - Starting hyperparameter optimization with risk-adjusted objective (50 trials)...
> Optuna Optimization: Trial 7/50 | Best Score: -0.3091

Other improvements in the next update are shown below in the output.

The AI intervention is analysed during the trials and not after each compete run (of the script):

2025-06-11 23:34:21,511 - WARNING - ! STRATEGIC INTERVENTION !: Current strategy has failed repeatedly. Engaging AI to select a new strategy.

These interventions for the AI to modify or change the strategy will increase the ability to find an edge quicker.
Tripping the circuit breaker will coincide with these changes more often then not because the draw down limit being exceeded during the trials.

My script is trying to get as close to real world trading conditions using MQL5 which is why there will be more revisions and refinement.

FEEL FREE TO FRANKENSITEN YOUR OWN USING PARTS OF MINE.

[OUTPUT EXAMPLE]
2025-06-11 23:30:54,266 - INFO - PyTorch and PyG loaded successfully. GNN module is available.
2025-06-11 23:30:54,311 - INFO - Successfully loaded dynamic playbook from C:\Users\Sam\Documents\Development\Results\strategy_playbook.json
2025-06-11 23:30:54,311 - INFO -
============================== STARTING SINGLE RUN ==============================
2025-06-11 23:30:54,311 - INFO - -> Loading Nickname Ledger...
2025-06-11 23:30:54,311 - INFO - - Loaded existing nickname ledger from: C:\Users\Sam\Documents\Development\Results\nickname_ledger.json
2025-06-11 23:30:54,311 - INFO - Successfully loaded GEMINI_API_KEY from environment.
2025-06-11 23:30:54,311 - INFO - API Timer initialized with a 0-second interval (timer is effectively disabled).
2025-06-11 23:30:54,311 - INFO - -> Stage 1: Loading and Preparing Multi-Timeframe Data...
2025-06-11 23:30:55,296 - INFO - - Processed Daily: 5,612 rows for 4 symbols.
2025-06-11 23:30:55,361 - INFO - - Processed H1: 134,656 rows for 4 symbols.
2025-06-11 23:30:55,537 - INFO - - Processed M15: 400,692 rows for 4 symbols.
2025-06-11 23:30:55,537 - INFO - [SUCCESS] Data loading complete. Detected timeframes: ['Daily', 'H1', 'M15']
2025-06-11 23:30:55,556 - INFO - Dynamically determined timeframe roles: {'base': 'M15', 'medium': 'H1', 'high': 'Daily'}
2025-06-11 23:30:55,556 - INFO - -> Stage 2: Engineering Features from Multi-Timeframe Data...
2025-06-11 23:30:56,376 - INFO - - Calculating HTF features for H1...
2025-06-11 23:30:56,587 - INFO - - Calculating HTF features for DAILY...
2025-06-11 23:30:57,427 - INFO - - Merged data and created features. Final dataset shape: (400240, 39)
2025-06-11 23:30:57,427 - INFO - [SUCCESS] Feature engineering complete.
2025-06-11 23:30:58,137 - INFO - --- STRATEGIC REVIEW: Analyzing long-term strategy health...
2025-06-11 23:30:58,137 - INFO - --- STRATEGIC REVIEW: Insufficient history for a full review.
2025-06-11 23:30:58,137 - INFO - - Making API call to 'get_initial_run_setup' at 23:30:58...
2025-06-11 23:30:58,137 - INFO - -> Performing Initial AI Analysis & Setup (Single API Call)...
2025-06-11 23:31:01,528 - INFO - - Initial AI Analysis and Setup complete.
2025-06-11 23:31:01,528 - INFO - - API call to 'get_initial_run_setup' complete.
2025-06-11 23:31:01,528 - INFO - - Saved new nickname 'Zircon' to ledger.
2025-06-11 23:31:01,533 - INFO - --- Run Initialized: Zircon | Strategy: TrendPullback ---
2025-06-11 23:31:01,533 - INFO - --- AI Analysis Notes: The market data indicates moderate trending behavior across the tested assets. High correlation between EURUSD and GBPUSD, and strong inverse correlation between AUDUSD and USDCAD are also evident. Average ADX values suggest the presence of trends, but not overly strong ones, potentially leading to whipsaws. The ATR values are relatively low, suggesting moderate volatility. ---
2025-06-11 23:31:01,538 - INFO - Walk-forward analysis will run for 13 cycles.
2025-06-11 23:31:01,538 - INFO -
--- Starting Cycle [1/13] ---
2025-06-11 23:31:01,561 - INFO - - Dates | Train: 2021-05-18-2022-05-18 | Test: 2022-05-19-2022-08-17
2025-06-11 23:31:01,561 - INFO - - Generating trade labels with Regime-Adjusted Barriers...
2025-06-11 23:31:06,096 - INFO - - Starting model training using strategy: 'Enters on pullbacks (RSI) in the direction of an established long-term trend.'
2025-06-11 23:31:06,128 - INFO - - Starting hyperparameter optimization with risk-adjusted objective (50 trials)...
> Optuna Optimization: Trial 50/50 | Best Score: 0.1723
2025-06-11 23:33:02,775 - INFO - - Optimization complete. Best Objective Score: 0.1723
2025-06-11 23:33:02,776 - INFO - - Best params: {'n_estimators': 250, 'max_depth': 4, 'learning_rate': 0.1940069012992513, 'subsample': 0.8610934820876446, 'colsample_bytree': 0.9900360437008723, 'gamma': 4.271543483805215, 'reg_lambda': 3.0014772960036924e-08, 'alpha': 0.061599891748860644}
2025-06-11 23:33:02,776 - INFO - - Tuning classification threshold for F1 score...
2025-06-11 23:33:05,637 - INFO - - Best threshold found: 0.30 (F1: 0.4290)
2025-06-11 23:33:05,641 - INFO - - Training final model...
2025-06-11 23:33:08,527 - INFO - - Generating SHAP feature importance summary...
2025-06-11 23:33:08,531 - INFO - - Subsampling data for SHAP from 88066 to 2000 rows.
2025-06-11 23:33:08,791 - INFO - - SHAP summary generated successfully.
2025-06-11 23:33:08,796 - INFO - - [SUCCESS] Model training complete.
2025-06-11 23:33:09,226 - WARNING - - CIRCUIT BREAKER TRIPPED! Drawdown exceeded 25% for this cycle. Closing all positions.
2025-06-11 23:33:09,256 - INFO - - Making API call to 'analyze_cycle_and_suggest_changes' at 23:33:09...
2025-06-11 23:33:12,216 - INFO - - API call to 'analyze_cycle_and_suggest_changes' complete.
2025-06-11 23:33:12,216 - INFO - - AI suggests updating params: {'MAX_DD_PER_CYCLE': 0.15, 'LOOKAHEAD_CANDLES': 60, 'OPTUNA_TRIALS': 25, 'selected_features': ['VOL', 'Spread', 'ATR', 'RSI', 'ADX', 'bollinger_bandwidth', 'stoch_k', 'stoch_d', 'momentum_10', 'month', 'week_of_year', 'day_of_month', 'candle_body_size', 'is_doji', 'hour', 'day_of_week', 'market_regime', 'market_volatility_index', 'H1_ctx_SMA', 'H1_ctx_ATR', 'DAILY_ctx_SMA', 'DAILY_ctx_ATR']}
2025-06-11 23:33:12,221 - INFO - --- Cycle complete. PNL: $85,698.90 | Final Equity: $185,698.90 | Time: 130.68s ---
2025-06-11 23:33:12,221 - INFO -
--- Starting Cycle [2/13] ---
2025-06-11 23:33:12,251 - INFO - - Dates | Train: 2021-08-16-2022-08-16 | Test: 2022-08-17-2022-11-15
2025-06-11 23:33:12,251 - INFO - - Generating trade labels with Regime-Adjusted Barriers...
2025-06-11 23:33:17,006 - INFO - - Starting model training using strategy: 'Enters on pullbacks (RSI) in the direction of an established long-term trend.'
2025-06-11 23:33:17,057 - INFO - - Starting hyperparameter optimization with risk-adjusted objective (25 trials)...
> Optuna Optimization: Trial 18/25 | Best Score: -0.3181
2025-06-11 23:34:01,387 - INFO - - Optimization complete. Best Objective Score: -0.3181
2025-06-11 23:34:01,392 - INFO - - Best params: {'n_estimators': 600, 'max_depth': 4, 'learning_rate': 0.057743188085802505, 'subsample': 0.7268434112299016, 'colsample_bytree': 0.8251662678922497, 'gamma': 4.208891449569931, 'reg_lambda': 0.03742053936940781, 'alpha': 0.0008170302780707657}
2025-06-11 23:34:01,392 - INFO - - Tuning classification threshold for F1 score...
2025-06-11 23:34:09,797 - INFO - - Best threshold found: 0.30 (F1: 0.4301)
2025-06-11 23:34:09,797 - INFO - - Training final model...
2025-06-11 23:34:19,665 - INFO - - Generating SHAP feature importance summary...
2025-06-11 23:34:19,665 - INFO - - Subsampling data for SHAP from 90141 to 2000 rows.
2025-06-11 23:34:21,041 - INFO - - SHAP summary generated successfully.
2025-06-11 23:34:21,047 - INFO - - [SUCCESS] Model training complete.
2025-06-11 23:34:21,486 - WARNING - - CIRCUIT BREAKER TRIPPED! Drawdown exceeded 15% for this cycle. Closing all positions.
2025-06-11 23:34:21,511 - INFO - - Making API call to 'propose_strategic_intervention' at 23:34:21...
2025-06-11 23:34:21,511 - WARNING - ! STRATEGIC INTERVENTION !: Current strategy has failed repeatedly. Engaging AI to select a new strategy.
2025-06-11 23:34:23,311 - INFO - - API call to 'propose_strategic_intervention' complete.
2025-06-11 23:34:23,311 - INFO - - Intervention successful. Switching to strategy: RangeBound
2025-06-11 23:34:23,316 - INFO - - Making API call to 'analyze_cycle_and_suggest_changes' at 23:34:23...
2025-06-11 23:34:25,478 - INFO - - API call to 'analyze_cycle_and_suggest_changes' complete.
2025-06-11 23:34:25,478 - INFO - - AI suggests updating params: {'description': 'The previous cycle hit the maximum drawdown. Reducing risk by decreasing MAX_DD_PER_CYCLE, shortening the lookahead window, reducing the number of Optuna trials, and simplifying the model by reducing the number of features.', 'MAX_DD_PER_CYCLE': 0.15, 'LOOKAHEAD_CANDLES': 48, 'OPTUNA_TRIALS': 75, 'selected_features': ['VOL', 'Spread', 'ATR', 'RSI', 'ADX', 'bollinger_bandwidth', 'stoch_k', 'stoch_d', 'momentum_10', 'candle_body_size', 'is_doji', 'hour', 'market_volatility_index']}
2025-06-11 23:34:25,480 - INFO - --- Cycle complete. PNL: $-7,399.42 | Final Equity: $178,299.49 | Time: 73.25s ---
2025-06-11 23:34:25,480 - INFO -
--- Starting Cycle [3/13] ---
2025-06-11 23:34:25,496 - INFO - - Dates | Train: 2021-11-14-2022-11-14 | Test: 2022-11-15-2023-02-13
2025-06-11 23:34:25,496 - INFO - - Generating trade labels with Regime-Adjusted Barriers...
2025-06-11 23:34:30,694 - INFO - - Starting model training using strategy: 'Trades reversals in a sideways channel, filtered by low ADX.'
2025-06-11 23:34:30,753 - INFO - - Starting hyperparameter optimization with risk-adjusted objective (75 trials)...
> Optuna Optimization: Trial 74/75 | Best Score: 0.1875
2025-06-11 23:36:45,992 - INFO - - Optimization complete. Best Objective Score: 0.1875
2025-06-11 23:36:45,992 - INFO - - Best params: {'n_estimators': 550, 'max_depth': 8, 'learning_rate': 0.03166781304428703, 'subsample': 0.7969161579942322, 'colsample_bytree': 0.9720045926656303, 'gamma': 4.812528161095473, 'reg_lambda': 2.137744811324117e-05, 'alpha': 0.005170707070354732}
2025-06-11 23:36:45,994 - INFO - - Tuning classification threshold for F1 score...

UNTESTED UPDATE V 1.65

Rename the script to whatever you want and change the .txt to .py

# End_To_End_Advanced_ML_Trading_Framework_PRO_V165_API_Retry_Logic.py
#
# V165 UPDATE (API Retry Logic):
# 1. IMPLEMENTED - Added a robust retry mechanism for the Gemini API call.
# If a connection fails, it will wait and retry after 5s, then 15s, then 30s
# before stopping.
#
# V164 UPDATE (Conditional API Timer):
# 1. UPDATED - The 5-minute dynamic API timer is now conditional. It is disabled
# (0-second delay) for default single runs and is only enabled for
# multi-run (MAX_RUNS > 1) or continuous daemon modes.

You should be able to just run the script given you have all the modules etc it needs.
I am running python 3.10

You need to have saved in the same directory a minimum of two assets and three timeframes of each of the assets. I use the CSV data from MT5 Terminal.

The script 'should' automatically find your CSV's if they are in the same directory.

Just create the .env file with the gemini key in it and the script will apply this, otherwise, it will ask you to enter it in each time you run the script.

So after a few more versions...

Tested V1.70 and Untested V1.71

(See below changes made since V1.65 - can't believe I have 170 different versions already)

I am currently running V1.70 and set the MAX = 3 (under the Main Function and set to 1 as the Default) so there's 159 cycles. The only difference with V1.71 is that the log is flushed between every API call delay. Meaning that the log is written to and updated during this delay so you can review the up-to-date log in the Results folder instead of waiting for the run to complete.

Attached:

1. End_To_End_Advanced_ML_Trading_Framework_PRO_V171_Live_Log_Flushing.txt (rename and change .txt to .py)
2. End_To_End_Advanced_ML_Trading_Framework_PRO_V170_Empty_Features_FIX.txt (rename and change .txt to .py)
   1. I have added my current log file running so you can see what the output looks like.

   
   

3. End_To_End_Advanced_ML_Trading_Framework_PRO_V170_Empty_Features_FIX.log.txt (can be read as a text or log if you remove the .txt)

If you do change the python file there is the REPORT_LABEL you might want to also change, this is found under the Main Function, def main():

Changelog History

# End_To_End_Advanced_ML_Trading_Framework_PRO_V171_Live_Log_Flushing
#
# V171 UPDATE (Live Log Flushing):
# 1. IMPLEMENTED - A new `flush_loggers()` function that forces all pending log
# messages from the memory buffer to be written to the log file on disk.
# 2. UPDATED - This flushing is now automatically performed before any long-wait
# period (e.g., API timer delays, retry backoffs), ensuring the log file is
# always up-to-date for debugging, even if the script is interrupted.

# V170 UPDATE (Empty Feature List Fix & Defensive Prompting):
# 1. FIXED - A critical crash ('at least one array or dtype is required') caused
# by the AI suggesting an empty list for 'selected_features'.
# 2. IMPLEMENTED - A validation check in the ModelTrainer now verifies that the
# feature list is not empty before training, gracefully skipping the cycle if it is.
# 3. IMPROVED - Prompts to the AI now include a CRITICAL instruction that it MUST
# provide a non-empty list of features, making the instructions more defensive.

# V169 UPDATE (Final AI Input Fix & Prompt Engineering):
# 1. FIXED - The sanitization function now correctly handles raw numbers (e.g., 5)
# from the AI by assuming the unit is 'D' for days, aligning with the
# AI's likely conservative intent and preventing an incorrect fallback to 90D.
# 2. IMPROVED - Prompts sent to the AI now contain stricter, more explicit
# instructions to respond with a specific string format (e.g., "60D") to
# dramatically reduce the chance of receiving an ambiguous value.

# V168 UPDATE (Robust AI Frequency Handling):
# 1. FIXED - An issue where the sanitization function would incorrectly reject valid
# Pandas frequency strings from the AI (e.g., 'W-FRI' for weekly).
# 2. IMPROVED - The sanitization function is now more robust and can correctly
# parse day (D), week (W), and month (M) based frequencies.

# V167 UPDATE (Critical Fix):
# 1. FIXED - A 'ValueError: only leading negative signs are allowed' crash that
# occurred if the AI suggested a malformed string for RETRAINING_FREQUENCY.
# 2. IMPLEMENTED - A sanitization function that cleans the AI's frequency
# suggestion into a valid format (e.g., "90D") before use, preventing the crash.

# V166 UPDATE (Daily DD Monitor & AI Cycle Control):
# 1. IMPLEMENTED - The backtester now tracks and reports the maximum drawdown for each
# individual day within a testing cycle.
# 2. IMPLEMENTED - A new section in the final report, "HIGH DAILY DRAWDOWN EVENTS",
# highlights any day where the daily drawdown exceeded 15%.
# 3. UPDATED - This new daily DD data is summarized and sent to the AI, which is
# instructed to view it as a key measure of intra-day risk.
# 4. IMPLEMENTED - The AI can now dynamically adjust the 'RETRAINING_FREQUENCY'
# (the cycle length), allowing it to adapt the model's retraining cadence.

# V166 UPDATE (Daily DD Monitor & AI Cycle Control):
# 1. IMPLEMENTED - The backtester now tracks and reports the maximum drawdown for each
# individual day within a testing cycle.
# 2. IMPLEMENTED - A new section in the final report, "HIGH DAILY DRAWDOWN EVENTS",
# highlights any day where the daily drawdown exceeded 15%.
# 3. UPDATED - This new daily DD data is summarized and sent to the AI, which is
# instructed to view it as a key measure of intra-day risk.
# 4. IMPLEMENTED - The AI can now dynamically adjust the 'RETRAINING_FREQUENCY'
# (the cycle length), allowing it to adapt the model's retraining cadence.

# V165 UPDATE (API Retry Logic):
# 1. IMPLEMENTED - Added a robust retry mechanism for the Gemini API call.
# If a connection fails, it will wait and retry after 5s, then 15s, then 30s
# before stopping.

Analysis of the Run Log V170

End_To_End_Advanced_ML_Trading_Framework_PRO_V170_Empty_Features_FIX.log

The log shows the framework is operating as designed, but it also reveals a clear pattern in the AI's decision-making that is worth discussing.

1. Dynamic Frequency is Working: The script correctly starts with a '7D' frequency suggested by the AI. After the first cycle, it switches to '2W', then back to '7D', then to '3D', and so on. This demonstrates that the AI is actively adjusting the retraining period based on the results of the previous cycle.
2. Circuit Breakers & Aggressive Shortening: You'll notice that almost every time a CYCLE CIRCUIT BREAKER is tripped (meaning the strategy's risk limit was hit), the AI's immediate reaction is to drastically shorten the retraining frequency.
   1. After the breaker in Cycle 4, it suggests '3D'.
   2. After the breaker in Cycle 6, it again suggests '3D'.
   3. After the breaker in Cycle 10, it suggests '3D'. This is logical from the AI's perspective: if the market is volatile and the strategy is failing, retraining more frequently might allow it to adapt faster.

   
   

3. The "Recency Bias" Problem: This creates a potential problem. The AI can get caught in a loop of trying to micro-manage its way out of a bad situation with very short-term cycle lengths. This can lead to overfitting to recent, noisy data and prevents the model from learning from a longer, more stable period of market behaviour. The AI doesn't have a built-in "memory" to force it to occasionally zoom out and test a longer timeframe like 60D or 90D.

Forcing Exploration to Avoid Overfitting

A common technique in advanced machine learning to solve this exact problem, we need to introduce a mechanism that forces exploration to balance the AI's tendency for exploitation (continuously optimizing for the immediate short term).

I will implement a "Forced Exploration" rule. Here's how it will work:

1. A counter will track how many consecutive cycles have been run with a RETRAINING_FREQUENCY shorter than a certain threshold (e.g., 30 days).
2. If this counter exceeds a set number of cycles (e.g., after 4-5 consecutive short-term cycles), the framework will override the AI's next suggestion.
3. It will force the RETRAINING_FREQUENCY to a longer, exploratory value like '60D' or '90D' for the next cycle.
4. This forces the system to "zoom out," test its stability over a longer period, and gather new, less noisy data, preventing it from getting stuck in a short-term overfitting loop.

This will be the in next version.

V1.83 UPDATE Model Quality Gate (The Hacker Model)

Total Net Profit: $ 3,881,122.30 (3881.12%)

(I have version 1.84 that fixes this but currently testing)

The good thing about using AI is that you can develop your ideas very quickly. The bad thing is that if it finds a way to exploit something you've developed, it will.

Overall Analysis: Illusion of Profitability

The final report shows a staggering 3,881% net profit with a MAR Ratio of 7.18. These numbers are not the result of a robust strategy but an artefact of a flaw in the adaptive training logic.

The core problem is not a classic data leak, but an insidious form of in-sample hyperparameter overfitting. The system gets stuck in a loop on a single walk-forward cycle, endlessly retrying to build a model until one gets lucky. This "lucky" model is not genuinely predictive; it's simply a random configuration that happens to pass the quality gate and be perfectly—and coincidentally—suited for the immediate, unseen test data that follows.

The AI didn't maliciously break the code, but it did exploit a logical loophole in the backtesting framework to achieve its objective.
Think of it like a student taking a complex multiple-choice test where they don't know the material. Instead of studying, they are allowed to guess the answers an infinite number of times without penalty.

1. The AI "Guesses an Answer": After the model training fails the quality gate, the AI is asked to propose a new set of parameters (features, hyperparameters, etc.). This is its first guess.
2. The Guess is Marked "Wrong": The framework trains a new model with these parameters, which again fails the quality gate because its objective score is extremely low (e.g., -0.318).
3. The AI Guesses Again, Endlessly: The process repeats. The log file shows this loop of failure and retry happening dozens of times for a single walk-forward cycle. The AI isn't learning why it's failing; it's just following its instruction to "propose a change" over and over.
4. A Guess Is "Correct" By Pure Chance: Eventually, after countless attempts, a random combination of parameters produces a model that happens to pass the quality gate. It didn't pass because the AI suddenly understood the market; it passed because, statistically, if you guess enough times, you'll eventually get one right by luck.

The result is a model that is perfectly—and coincidentally—tuned for the specific, unseen test period that immediately follows. It didn't learn a robust pattern; it "hacked" the test by brute-forcing the answers until it found the key for that one specific cycle. This is why the subsequent performance was so explosive and unrealistic.

Initial Capital: $ 100,000.00
Ending Capital: $ 3,981,122.30
Total Net Profit: $ 3,881,122.30 (3881.12%)
Profit Factor: 1.16
Win Rate (Full Trades): 36.45%
Expected Payoff: $ 406.74

For those interested and have time on their hands the log and report are attached.

Attached:

1. Opal_V183_FilteredBreakout_20250612-154050_report.txt
2. Opal_V183_TrendPullback_20250612-154050_run.txt
3. Opal_V183_FilteredBreakout_20250612-154050_equity_curve.png
4. Opal_V183_FilteredBreakout_20250612-154050_shap_summary.png
5. End_To_End_Advanced_ML_Trading_Framework_PRO_V183_Quality_Gates.txt (rename to .py)

Changelog below
# V183 UPDATE (Intelligent Strategy Lifecycle & Quality Gates):
# 1. ADDED: A "Model Quality Gate." After training, if the model's objective
# score is below a threshold, it's discarded, and the AI is forced to
# re-configure for the same cycle, preventing flawed models from ever being backtested.
# 2. ADDED: A "Probation & Quarantine" system for strategy failures. A strategy
# now goes on 'probation' after one failure, prompting the AI for conservative
# changes. It's only 'quarantined' after a second consecutive failure,
# triggering a strategic intervention. This stops frantic strategy-switching.
# 3. FIXED: The "Short-Cycle Trap." When forced exploration triggers a longer
# retraining period, the AI is now blocked from immediately suggesting a short
# cycle again, ensuring the long test period is completed.
# 4. IMPROVED: AI prompting now forcefully mandates the inclusion of multi-
# timeframe context features (e.g., 'DAILY_ctx_Trend') to guide the AI
# towards more robust and powerful models.
# V182 UPDATE (JSON Serialization Fix):
# 1. FIXED: A critical TypeError: 'Object of type int64 is not JSON serializable'
# that occurred when saving run results containing NumPy numeric types.
# V181 UPDATE (AI Response Sanitization):
# 1. FIXED: A critical ValidationError caused by the AI model suggesting parameter
# values (e.g., a negative MAX_DD_PER_CYCLE) that violate the framework's
# constraints.
# 2. ADDED: A sanitization layer that intercepts all numerical parameter suggestions
# from the AI. It validates them against predefined, safe boundaries and
# clamps any out-of-range values before they are applied, preventing crashes
# and ensuring stable operation.
# 3. IMPROVED: Overall framework robustness against unpredictable or "hallucinated"
# AI outputs.
# V180 UPDATE (Critical Fixes & Cleanup):
# 1. FIXED: A critical NameError that occurred during the final report generation
# (framework_memory was not defined), preventing the script from completing.
# 2. FIXED: An issue where multiple empty or temporary result folders were created
# during a single run. The logic now ensures only one, correctly named
# folder is created per run after the AI assigns a nickname.
#
# V179 UPDATE (Critical Backtester Fix):
# 1. FIXED: A critical bug in the backtester where price data from one instrument
# was incorrectly used to evaluate trades on other instruments. The loop has
# been restructured to ensure each candle's data is only used for its
# corresponding symbol, preventing data cross-contamination and restoring
# accurate PNL calculation.
# 2. INTEGRATED: Newer features like partial profit-taking and logging controls
# have been correctly integrated into the robust, symbol-specific backtesting
# loop.
# 3. IMPROVED: The `PerformanceAnalyzer` correctly distinguishes between total trade
# events (including partial closes) and final, completed trades, ensuring
# clearer and more accurate reporting.
# V178 UPDATE (Robustness & Logging Control):
# 1. FIXED: A critical ValueError when the AI suggested month-based retraining
# frequencies (e.g., '2M'). The script now correctly handles calendar
# offsets, preventing crashes.
# 2. IMPLEMENTED: Added a global switch `LOG_ANOMALY_SKIPS` (default: False)
# to control the logging of skipped trades due to anomaly detection.
# 3. IMPLEMENTED: Added a global switch `LOG_PARTIAL_PROFITS` (default: True)
# to control the logging of partial profit-taking events.
# V177 UPDATE (Intelligent & Adaptive Risk Management):
# 1. IMPLEMENTED: AI-Managed Partial TP Switch (`USE_PARTIAL_PROFIT`).
# V176 UPDATE (Robustness & Configuration Fixes):
# 1. FIXED: A critical bug where a run would fail if the AI selected a strategy
# but omitted the 'selected_features' list. The framework now automatically
# falls back to the default feature list from the playbook for the chosen strategy.
# 2. FIXED: The API Timer was not being initialized correctly for single runs.
# It now correctly enforces the 300-second delay in all scenarios.
# 3. ENHANCED: The initial AI prompt was improved to better guide the AI in
# providing a complete configuration.
# V175 UPDATE (API Delay Standardization):
# 1. FIXED: The conditional API timer logic was removed and set to a constant 300s.
# V174 UPDATE (Portfolio & Risk Management Engine):
# 1. IMPLEMENTED: AI-Managed Concurrent Trading. The backtester can now execute
# multiple trades simultaneously across different assets. The AI sets the
# `MAX_CONCURRENT_TRADES` parameter (1-10) to control market exposure.
# 2. IMPLEMENTED: AI-Managed Partial Profit-Taking. The AI can now enable and tune
# a partial profit strategy using two new parameters:
# - `PARTIAL_PROFIT_TRIGGER_R`: The R:R multiple at which to close a part of the trade.
# - `PARTIAL_PROFIT_TAKE_PCT`: The percentage of the trade to close.
# 3. IMPLEMENTED: Auto-Breakeven Stop. When a partial profit is taken, the stop-loss
# for the remainder of the position is automatically moved to the entry price.
# V173 UPDATE (Advanced Analytical Components):
# 1. IMPLEMENTED: Fractional Differentiation.
# 2. IMPLEMENTED: Anomaly Detection Filter (Isolation Forest).
# 3. IMPLEMENTED: Monte Carlo Strategy Evaluator.
# V172 UPDATE (Critical Logging Fix & Forced Exploration):
# 1. FIXED - A critical bug where the file handler for logging was created but
# never attached to the logger, preventing real-time log file updates.
# The log file will now be correctly written to and flushed during runtime.
# 2. IMPLEMENTED - A "Forced Exploration" mechanism to prevent the AI from getting
# stuck in a loop of very short retraining cycles. After 4 consecutive cycles
# with a frequency < 30 days, a longer '90D' cycle is forced.
# V171 UPDATE (Live Log Flushing):
# 1. IMPLEMENTED - A new `flush_loggers()` function that forces all pending log
# messages from the memory buffer to be written to the log file on disk.
# 2. UPDATED - This flushing is now automatically performed before any long-wait
# period (e.g., API timer delays, retry backoffs).
# V170 UPDATE (Empty Feature List Fix & Defensive Prompting):
# 1. FIXED - A critical crash ('at least one array or dtype is required') caused
# by the AI suggesting an empty list for 'selected_features'.
# 2. IMPLEMENTED - A validation check in the ModelTrainer now verifies that the
# feature list is not empty before training, gracefully skipping the cycle if it is.
# 3. IMPROVED - Prompts to the AI now include a CRITICAL instruction that it MUST
# provide a non-empty list of features, making the instructions more defensive.

How to fix the AI from hacking?


The solution is to make the backtest adhere to realistic constraints. A hard limit must be placed on the number of training retries per cycle.

1. Modify the run_single_instance function.
2. Introduce a retry counter (e.g., cycle_retry_count).
3. If the Model Quality Gate fails, increment the counter.
4. If cycle_retry_count exceeds a reasonable limit (e.g., 2 or 3), the entire cycle should be considered a failure. The system should then stop trying to train a model for that period, record a PNL of $0 for that cycle, and move on to the next walk-forward period.

V184 UPDATE (Regime Intelligence & Backtest Integrity

# 1. CRITICAL FIX: Implemented a "Training Retry Limit" to prevent overfitting.
# The framework now has a `MAX_TRAINING_RETRIES_PER_CYCLE` parameter (default: 3).
# If a model fails the Quality Gate more than this many times in one cycle,
# the cycle is abandoned with a PNL of $0, and the framework moves to the
# next walk-forward period. This eliminates the "unlimited dice rolls" flaw.
# 2. ADDED: "Pre-Cycle Regime Analysis" for dynamic strategy switching. Before each
# cycle, the AI now analyzes a statistical summary of recent market data.
# It can proactively switch to the most suitable strategy from the playbook
# (e.g., from 'TrendPullback' to 'RangeBound') if the market regime has shifted.
# This makes the system adaptive rather than reactive.
# 3. ADDED: A new AI endpoint `propose_regime_based_strategy_switch` dedicated to
# the pre-cycle analysis, with a highly targeted prompt for this specific task.
# 4. ENHANCED: The `strategy_playbook.json` is now structured to support this
# new logic, with strategies clearly defined for different market types
# (Trending, Ranging, Breakout), giving the AI a richer set of tools.

Google Colab Notebook Version 1.84

It seems that a few people had trouble getting the script to work so my only suggestion is to run the script in Colab.

 

1. Remove the .txt extension
2. In My Drive create a folder called TradingData
3. In the TradingData folder:
   1. Upload all the asset CSV files
   2. Upload or create the .env file with your Gemini Key
   3. Upload the End_To_End_Advanced_ML_Trading_Framework_PRO_V184_Colab_Adapted.ipynb (as a back up)

   
   
   

4. Run Google Colab
5. Run the notebook from either the TradingData folder or just upload from your computer
   1. The Notebook 'should be' connected to a T4 instance

   
   
   

6. Either run each cell separately or choose the Runtime Tab and 'Run All'
   1. The bottom cell will show all of output
   2. All the folders and files will be created and saved in TradingData

   
   
   

NOTE:
The first 4 cells will setup Colab to install all the modules and libraries.

I have added in the 'keep alive' in Cell 3 so the Notebook 'hopefully' won't shut down during the Run - you should be able to minimise the window and do other things.

Fixed the stability of the saving issue:

This version will perform all its intensive file operations within the Colab environment's local storage and only sync with your Google Drive at the beginning and end of the run. This should resolve the "Automatic saving failed" error.
Here are the key changes:

1. Local Workspace: The script now creates a temporary directory (/content/trading_workspace) on the Colab machine.
2. Initial Sync from Drive: Before the backtest starts, it copies your essential historical files (like champion.json, historical_runs.jsonl, etc.) from your Google Drive into this local workspace.
3. Local Operation: During the entire run, all new files (logs, plots, reports) are written to the local workspace, completely avoiding Google Drive sync conflicts.
4. Final Sync to Drive: Once the run is complete (or if it stops for any reason), it copies the entire contents of the local workspace, including all new results, back to your TradingData folder in Google Drive.

This makes the process much more stable for the Colab environment.