Inverse Watch Docs
AppLanding
  • Overview
    • Home
    • Governance
      • Proposal 7
      • Proposal 25
      • Proposal 52
      • Proposal 107
      • Proposal 147 - S1
      • Proposal 189 - S2
  • Products
    • Inverse Alerts
      • See on Twitter
    • Inverse Chatbot
      • /doc
      • /imagine
      • /data
      • /graph
    • Inverse Subgraphs
      • See inverse-subgraph on Mainnet
      • See inverse-governance-subgraph on Mainnet
    • Inverse Watch
      • Go to App
  • User Guide
    • Quickstart
    • Alerts
      • Setting Up an Alert
      • Adding New Alert Destinations
      • Customize Alert Template
      • Multiple Column Alert
    • Queries
      • Creating and Editing Queries
      • Querying Existing Query Results
      • Query Parameters
      • How to Schedule a Query
      • Favorites & Tagging
      • Query Filters
      • How To Download / Export Query Results
      • Query Snippets
    • Visualizations
      • Cohort Visualizations
      • Visualizations How-To
      • Chart Visualizations
      • Formatting Numbers in Visualizations
      • How to Make a Pivot Table
      • Funnel Visualizations
      • Table Visualization Options
      • Visualizations Types
    • Dashboards
      • Creating and Editing Dashboards
      • Favorites & Tagging
      • Sharing and Embedding Dashboards
    • Data Sources
      • CSV & Excel Files
      • Google Sheets
      • JSON (API)
      • Python
      • EVM Chain Logs
      • EVM Chain State
      • GraphQL
      • Dune API
    • Machine Learning
      • Data Engineering
      • Regressors
        • Linear Regression
        • Random Forest
        • Ada Boosting
        • Gradient Boosting
        • Neural Network (LSTM)
      • Training and Predicting
      • Metrics & Overfitting
      • Examples
        • Price Prediction
          • Data Preprocessing
          • Model Creation & Training
          • Metrics Evaluation
          • Back Testing
          • Visualizing
        • Liquidation Risk
  • Admin & Dev Guide
    • Setup
    • Redash
    • Integrations & API
    • Query Runners
    • Users
      • Adding a Profile Picture
      • Authentication Options
      • Group Management
      • Inviting Users to Use Redash
      • Permissions & Groups
    • Visualizations
  • Cheat Sheets
    • Snippets
    • Contracts
  • More
    • Deprecated Apps
    • Github : inverse-flaskbot
    • Github : inverse-subgraph
    • Github : inverse-watch
Powered by GitBook
On this page
  • Raw data table
  • Raw Price Prediction
  • Log Price Returns
  • Predicted vs. Real Price Using Log Price Returns
  • Trading Simulation
  • PnL (Profit and Loss)
  • Balance Over Time
  • Cumulative Returns
  • Drawdown Analysis
  • Conclusion

Was this helpful?

  1. User Guide
  2. Machine Learning
  3. Examples
  4. Price Prediction

Visualizing

PreviousBack TestingNextLiquidation Risk

Last updated 6 months ago

Was this helpful?

Inverse Watch provides a set of powerful visualization tools, allowing us to analyze our models' performance over time. By leveraging various charts, we gain deeper insights into the predictions and understand the trading performance better. Our strategy focuses on predicting trade direction (whether the price will increase or decrease), which simplifies decision-making compared to predicting exact price levels.


Raw data table

This table shows detailed metrics for each model, including actual prices, predicted prices, trade signals, PnL, balance, cumulative returns, and draw downs. This granular data is crucial for in-depth analysis of each model's performance over time.


Raw Price Prediction

The Raw Price Prediction chart compares actual prices with predictions from various models (e.g., Linear Regression, Random Forest, Neural Network, Gradient Boosting, and AdaBoost).

Key Insights:

  • Extreme Predictions: All models show sharp overestimations, particularly during volatile periods.

  • Unrealistic Predictions: The Linear model predicts negative prices at times, which is not realistic in financial settings.

These issues demonstrate the need to switch to logarithmic prices, which help reduce variance and ensure a more stable prediction environment. By using log prices, we eliminate unrealistic values like negative prices and limit extreme fluctuations.

However, since our focus is on trade direction, the accuracy of the prediction trend (whether the price is going up or down) is more relevant than the absolute price prediction.


Log Price Returns

Switching to log price returns results in more stable predictions. In the chart below, we compare predicted log price changes with actual log price changes across models.

Key Insights:

  • Improved Stability: Using log prices reduces spikes in predictions and provides a more consistent trend alignment.

  • Accurate Trade Direction: Models follow the actual price direction more closely, improving the precision of trade signals.

This approach provides more reliable predictions of trade direction, which is key to making correct buy/sell decisions.


Predicted vs. Real Price Using Log Price Returns

After switching to log prices, we compare predicted prices based on log price changes to actual prices. This comparison helps gauge the overall improvement in predicting price direction, which is the core of our trading strategy.

Key Insights:

  • Improved Trade Direction Prediction: Most models align more closely with actual price movements, indicating improved directional accuracy.

  • Stable Performance: The Neural Network and Gradient Boosting models show particularly stable results compared to others.

This confirms that log price changes offer a more effective approach for predicting price direction rather than exact price levels.


Trading Simulation

With accurate predictions of trade direction in place, we simulate trading strategies based on these predictions. The following charts break down trading performance for each model using directional trades rather than exact price levels.


PnL (Profit and Loss)

This chart visualizes the profit or loss generated by each model’s directional predictions.

Key Observations:

  • Higher Volatility: Random Forest and Gradient Boosting models generate larger spikes in PnL, indicating high volatility in directional predictions.

  • Steady Returns: Neural Networks and Linear Regression models show more stable PnL growth based on directional trades.


Balance Over Time

This chart shows how the account balance evolves for each model throughout the trading simulation, based on the accuracy of trade direction.

Key Observations:

  • Aggressive Growth: Random Forest exhibits the highest balance growth but carries increased risk due to more volatile directional predictions.

  • Steady Growth: Neural Network and Gradient Boosting models show more consistent balance growth, which is important for risk-averse traders.


Cumulative Returns

Cumulative returns show the accumulated profit or loss over time for each model, based on correct directional trades.

Key Observations:

  • High Returns with Risk: The Random Forest model shows the highest cumulative return but is also the most volatile in terms of trade direction.

  • Consistent Performance: Neural Network and Gradient Boosting models provide steadier cumulative returns based on more accurate directional predictions.


Drawdown Analysis

Drawdowns reflect the largest percentage decline in the account balance over time, helping assess the risk exposure for each model based on incorrect trade direction.

Key Observations:

  • Risky Models: Random Forest and AdaBoost experience large drawdowns due to incorrect trade direction, indicating higher risk exposure.

  • Controlled Risk: Neural Network and Gradient Boosting models maintain lower drawdown levels, making them safer choices for long-term trading strategies focused on directional accuracy.


Conclusion

The visualizations from Inverse Watch provide a comprehensive view of model performance and trading outcomes. Here are the key takeaways:

  • Prediction Accuracy: Log price changes offer more stable predictions, improving directional accuracy in real market movements.

  • Risk vs. Reward: Random Forest yields higher returns but comes with greater volatility and drawdowns. Meanwhile, Neural Network and Gradient Boosting models strike a better balance between directional accuracy, returns, and risk.

  • Long-term Performance: Cumulative returns and balance growth suggest that Neural Network and Gradient Boosting models are the most viable for risk-averse traders focused on trade direction.

By integrating these visualizations with statistical backtesting, traders can better understand the strengths and weaknesses of each model’s directional predictions, leading to more informed decision-making and enhanced trading strategies.

Warning !

Remember, while these visualizations provide valuable insights, they should be combined with rigorous statistical analysis and back testing before making any trading decisions. The performance in this historical data may not necessarily predict future performance.