> ## Content Index
> Fetch the complete content index at: https://www.hubwiz.com/blog/llms.txt
> Use this file to discover other available public pages before exploring further.

# TensorTrade 强化学习算法交易包
- URL: https://www.hubwiz.com/blog/tensortrade-rl-algo-trading-package/
- Published: 2025-11-16T15:45:39.000Z
- Updated: 2025-11-16T15:46:14.000Z
- Description: TensorTrade是一个基于强化学习的算法交易Python库，它是开源的，高度灵活，并且可以与我们常用的Python工具无缝协作
- Author: admin
- Tags: LIBRARY, STOCK TRADING, #Import 2026-10-14 14:41

我最近发现了一个名为TensorTrade的Python库。如果你曾经了解过强化学习，并且心想：“嗯……挺酷的，但还是别用了”，那么这个库会让你觉得整个主题不再那么令人生畏。它是开源的，高度灵活，并且可以与我们常用的Python工具无缝协作：NumPy、Pandas、Gym、TensorFlow、Keras……你应该明白我的意思了。

总之，让我带你了解一下我使用它时学到的东西。没什么特别的——只是足以让你了解各个部分是如何连接的。

## 1、获取价格数据

在深入学习强化学习之前，显然你需要一些历史数据。我从 yfinance 获取数据，因为它速度快，而且不需要签署 12 份不同的协议。

以下是我使用的完整代码：

```
import yfinance
import pandas_ta as ta

TICKER = 'TTRD'  # Replace this with whatever ticker you're actually trading
TRAIN_START_DATE = '2021-02-09'
TRAIN_END_DATE = '2021-09-30'
EVAL_START_DATE = '2021-10-01'
EVAL_END_DATE = '2021-11-12'
yf_ticker = yfinance.Ticker(ticker=TICKER)
# Training set
df_training = yf_ticker.history(start=TRAIN_START_DATE, end=TRAIN_END_DATE, interval='60m')
df_training.drop(['Dividends', 'Stock Splits'], axis=1, inplace=True)
df_training["Volume"] = df_training["Volume"].astype(int)
df_training.ta.log_return(append=True, length=16)
df_training.ta.rsi(append=True, length=14)
df_training.ta.macd(append=True, fast=12, slow=26)
df_training.to_csv('training.csv', index=False)
# Evaluation set
df_evaluation = yf_ticker.history(start=EVAL_START_DATE, end=EVAL_END_DATE, interval='60m')
df_evaluation.drop(['Dividends', 'Stock Splits'], axis=1, inplace=True)
df_evaluation["Volume"] = df_evaluation["Volume"].astype(int)
df_evaluation.ta.log_return(append=True, length=16)
df_evaluation.ta.rsi(append=True, length=14)
df_evaluation.ta.macd(append=True, fast=12, slow=26)
df_evaluation.to_csv('evaluation.csv', index=False)
```

这里没什么特别的——只是数据清理和一些指标，这样智能体就不会盲目交易。

## 2、构建环境

TensorTrade 的功能有点像简化的 Gym 环境。你提供价格数据，它会为你的智能体提供一个沙箱，让它从错误中学习。

以下是环境函数。别担心，它并不复杂：

```
import pandas as pd

from tensortrade.feed.core import DataFeed, Stream
from tensortrade.oms.instruments import Instrument
from tensortrade.oms.exchanges import Exchange, ExchangeOptions
from tensortrade.oms.services.execution.simulated import execute_order
from tensortrade.oms.wallets import Wallet, Portfolio
import tensortrade.env.default as default

def create_env(config):
    dataset = pd.read_csv(config["csv_filename"], parse_dates=['Datetime']).fillna(method='backfill')
    commission = 0.0035
    price = Stream.source(list(dataset["Close"]), dtype="float").rename("USD-TTRD")
    exchange_options = ExchangeOptions(commission=commission)
    ttse_exchange = Exchange("TTSE", service=execute_order, options=exchange_options)(price)
    USD = Instrument("USD", 2, "US Dollar")
    TTRD = Instrument("TTRD", 2, "TensorTrade Corp")
    cash = Wallet(ttse_exchange, 1000 * USD)
    asset = Wallet(ttse_exchange, TTRD)
    portfolio = Portfolio(USD, [cash, asset])
    renderer_feed = DataFeed([
        Stream.source(list(dataset["Datetime"])).rename("date"),
        Stream.source(list(dataset["Open"]), dtype="float").rename("open"),
        Stream.source(list(dataset["High"]), dtype="float").rename("high"),
        Stream.source(list(dataset["Low"]), dtype="float").rename("low"),
        Stream.source(list(dataset["Close"]), dtype="float").rename("close"),
        Stream.source(list(dataset["Volume"]), dtype="float").rename("volume")
    ])
    features = []
    for c in dataset.columns[1:]:
        features.append(
            Stream.source(list(dataset[c]), dtype="float").rename(c)
        )
    feed = DataFeed(features)
    feed.compile()
    reward_scheme = default.rewards.SimpleProfit(window_size=config["reward_window_size"])
    action_scheme = default.actions.BSH(cash=cash, asset=asset)
    env = default.create(
        feed=feed,
        portfolio=portfolio,
        action_scheme=action_scheme,
        reward_scheme=reward_scheme,
        renderer_feed=renderer_feed,
        renderer=[],
        window_size=config["window_size"],
        max_allowed_loss=config["max_allowed_loss"]
    )
    return env
```

目前，您实际上拥有一个小型交易环境，您的代理可以在其中买入、卖出或持有并等待最佳结果。

## 3、使用 Ray 运行训练

如果你喜欢自讨苦吃，可以手动训练强化学习智能体，但 Ray 简化了超参数调优过程。

以下是设置：

```
import ray
import os
from ray import tune
from ray.tune.registry import register_env

FC_SIZE = tune.grid_search([[256, 256], [1024], [128, 64, 32]])
LEARNING_RATE = tune.grid_search([0.001, 0.0005, 0.00001])
MINIBATCH_SIZE = tune.grid_search([5, 10, 20])
cwd = os.getcwd()
ray.init()
register_env("MyTrainingEnv", create_env)
env_config_training = {
    "window_size": 14,
    "reward_window_size": 7,
    "max_allowed_loss": 0.10,
    "csv_filename": os.path.join(cwd, 'training.csv'),
}
env_config_evaluation = {
    "max_allowed_loss": 1.00,
    "csv_filename": os.path.join(cwd, 'evaluation.csv'),
}
analysis = tune.run(
    run_or_experiment="PPO",
    name="MyExperiment1",
    metric="episode_reward_mean",
    mode="max",
    stop={"training_iteration": 5},
    config={
        "env": "MyTrainingEnv",
        "env_config": env_config_training,
        "framework": "torch",
        "num_workers": 1,
        "lr": LEARNING_RATE,
        "model": {"fcnet_hiddens": FC_SIZE},
        "sgd_minibatch_size": MINIBATCH_SIZE,
        "evaluation_interval": 1,
        "evaluation_config": {
            "env_config": env_config_evaluation,
            "explore": False
        }
    },
    num_samples=1,
    checkpoint_freq=1
)
```

Ray 本质上是尝试各种组合，多次运行你的环境，并告诉你哪种组合效果最佳，而无需你整晚坐在那里调整参数。

## 4、创建你自己的奖励函数

这部分让我很惊喜——TensorTrade 比大多数强化学习库都更简化了自定义奖励的创建过程。

以下是奖励类的示例：

```
from tensortrade.env.default.rewards import TensorTradeRewardScheme
from tensortrade.feed.core import DataFeed, Stream

class PBR(TensorTradeRewardScheme):
    registered_name = "pbr"
    def __init__(self, price: Stream):
        super().__init__()
        self.position = -1
        r = Stream.sensor(price, lambda p: p.value, dtype="float").diff()
        position = Stream.sensor(self, lambda rs: rs.position, dtype="float")
        reward = (r * position).fillna(0).rename("reward")
        self.feed = DataFeed([reward])
        self.feed.compile()
    def on_action(self, action: int):
        self.position = 1 if action == 0 else -1
    def get_reward(self, portfolio):
        return self.feed.next()["reward"]
    def reset(self):
        self.position = -1
        self.feed.reset()
```

没什么特别的——只是将价格变化与你的多空仓位关联起来。

## 5、结束语

TensorTrade 并不完美，没错，如果你的奖励函数很糟糕，你的智能体肯定会崩溃。但作为一个通过交易这种实际操作来学习强化学习的框架，它确实是我尝试过的最简单、最有趣的工具之一。

---

原文链接：[TensorTrade: Build AI-Driven Trading Algorithms with Reinforcement Learning](https://medium.com/coding-nexus/tensortrade-build-ai-driven-trading-algorithms-with-reinforcement-learning-27c77e96d7c6?ref=hubwiz.com)

汇智网翻译整理，转载请标明出处