- Create a gym environment for our agent to learn from (env.py)
- Render a simple, yet elegant visualization of that environment (trading_graph.py)
- Train our agent to learn a profitable strategy (train.py)
- Evaluate our agent's performance on test data and real-time trading (test.py)
- Environment: python 3.6
- Requirement: see requirements.txt
- Data source: https://www.histdata.com/download-free-forex-data/
- Time frame: 1 minutes (230300 <=> 23:03:00)
- Action space: buy, sell, hold, close, close then buy, close then sell.
- Observation space: open, close, high, low, agent's networth, usd held, eur held, actions in previous 60 time step
- Reward: networth(t) - networth(t-1)
- Win/lose percent of agent, our target is 70-80% (current rate: 50-60%)
- Risk to reward ratio, target is 1:4 (the best we've achieved so far is 1:3)
- Maximum drawdown <= 0.1% of current balance (done)
- Avg win value > 0.5% current balance or 40 pip is prefered
- Intergrate higher time frame
- Use pip instead of money to represent balance, reward...
- Research why testing result on training and testing data set is quite different (we expect it wouldn't vary too much)
- Reorganize codes (halfly done)
- Improve model accuracy (better reward function, more features as input to our data) (doing)
- Display trade history
- Make action space more diversity
- Write unit test
- Implement weight initialization
- Force all source codes to use relative file path while running
- Use heiken-ashi candle
- Normalize reward and other input features (done)
- Using news data as feature, we must calculate avg time the market is affected by news and then distribute it accordingly to current timeframe (done)
- Investigate why our win rate is 50% eventhough we did overfit the data (done, mostly because we train it incorrectly)
- Setup progressive training environment
- Add close orders action, immediately release all eur agent is holding or selling
- Split data to episodes, each episode contains 1 week data, shuffle those episodes
- Add more data (we are using data of year 2013 now)
- Create train and test data set
- Implement test procedure on test data set
- Data visualization
- Create some tools to compare performance between models
- Create config parser
- Overfit one batch of dataset to test model ability
- Implement save and restore model functionality
- Price changes dramatically in 15 mins before and 1h mins after critical news
- MT4 time zone: UTC+3, history calendar time zone: UTC-4, our data timezone: UTC-5
- Setup end-to-end training/evalution skeleton first + get dump baselines.
- Fix random seed, start simple first, don't add too many features to our model
- Complexity only one at a time (plug signal in one by one)
- M15 time frame data contains about 200k records, we may want to switch to smaller time frame for more data
- Right now we have reduced the spread to 0.1, in reality the spread will be 0.3 or higher,
we need to test it against higher spread - Currently, we don't have any benchmark system to tell us that our agent has perform the most correct action (achieve highest net worth possible) in those data frames
-
We can overfit mlp policy after 1m steps of training with data in a very short duration (300x15 minutes)
-
Eventhough we did overfit the model and it did increase our net worth overtime, but the win ratio is still 50%,
we may need to investigate later, because our goal is win rate >= 80% -
Win rate is not everything in Fx trading but we still need it higher 50% for more security.
-
Althought our model is overfitted is still perform prety well on test set: x6 networth in about 6000 steps,
It also go bankrupt in 1-2000 steps later on which is the result that we expected. -
Our objective is increase the avg win value per trade and win ratio as well as reduce the avg lose per trade
=> First conclusion: more data is not really good for model if we don't understand and arrange the data correctly, on the other hand, more training is really important in RL -
Agent sometimes buy or sell eur in agressive manner, it performs the same order in many continous steps. Which is why it loses or wins an immense amount of money, but it's not really what we want.
-
We would like the agent to control the risk-to-reward ratio that maximum drawdown is as low as possible
-
However, it means that we won't win big amount because we don't allow the agent to invest too much in one currency
-
This is some kind of trade-off and we rather play safe than put our capital at risk.
=> We need a reward function that punish agent for holding too much eur.
i.e. function that take into account the risk of producing high return
=> Our model learn to hold a numerous amount of eur in order to achieve great return.
However, this strategy is a double-edge sword, we could go bankrupt because of this
So, great rewards doesn't have to come from great returns, we could reward the agent when it wins small consecutive steps and not losing continously -
We have two kinds of agent at the moment: one has 50% win rate, avg win/lose value is 100/30 and the other win rate is 60%, avg win/lose is 30/10
=> This means agent has to sacrifice win rate to increase win/lose value.
- Small Reward for not losing - but how we define not losing??? if we lose 0 pip, it's not losing Reward: 0.01
- Medium Reward for winning small amount - how much winning is considered small amount??? 1 -> 5 pip is small amount Reward: 0.1
- Large Reward for winning consecutive steps: if an agent hold currency in a long time and the money flow keep going by it's current direction that's mean the risk we take is relatively small we should encourage our agent to do that How many steps is considered consecutive??? >= 2 steps Reward: 0.2 * steps + reward for amount of pip that it has winned
- Medium Punishment for holding too much currency: The more currency we hold the higher the risk we take, so we have to punish our agent by doing so More than 200k eur is considered pretty much Reward: -0.2 * (amount/100k) if amount is > 100k
- Small Punishment for losing small amount Reward: -0.1
- Medium Punishment for losing consecutive steps Reward: -0.2 * steps + rewar d for amount of pip that it has lost
- Consider the risk for leaving or enter a state (shape reward function) What happen if we close a trade??? if we close a trade with positive return => win trade => positive reward, otherwise would we punish our agent by closing a losing trade or let it continue to lose because it could become a winning trade in the future??? Reward: 0.2
- When should we close a trade? when we see the price is going against us or we don't know for sure which direction the price is going How do we reward this action??? Reward: 0.3 when close trade with positive return, penalize small reward (-0.2) when trade is closed with negative return
- Should we put average winning value into account? TBD
- Small lost is allowed but we must penalize big loss heavier than big win