🤖AI Forecaster

This guide will walk you through setting up, training, and using our AI Forecaster to make predictions which can then be used as inputs in your other modules.

Let’s start with the basics. What is forecasting you may ask?

In financial trading, forecasting involves machine learning algorithms that analyze historical financial data to identify patterns, trends, and market conditions. By understanding these past market behaviors, the algorithms can make data-driven predictions, or forecasts, about which direction prices are likely to move in the future.

Setting Up the Forecaster

Navigate to the ‘Modules’ tab and click on the ‘Add Module’ button. Next, select ‘Custom Module’ and ‘Build Your Own’. From the list of available modules, select the ‘AI Forecaster’.

Now, you can start setting up your Forecaster:

1. Basic Parameters Setup

Exchange: This is where you choose where your data comes from. Think of it like selecting a library to find your books. For example, if you pick "Binance," the forecaster will use data from the Binance library to understand and predict your selected cryptocurrency's behavior.

Symbol: Here, you decide the specific 'book' you want to study from our library. Choosing a trading pair like "BTC/USDT" tells the forecaster to focus on the story between Bitcoin and US Dollar Tether. Your model's predictions are like the summary of this book, so choose the one you're most interested in!

Prediction Period: Imagine deciding how often you want to take notes while reading your book. If you select "5m," you'll jot down a note every 5 minutes, giving you a detailed account. This helps determine how fine-grained your model's predictions are. The more frequent the notes, the more detailed the forecast, but also the more data you'll need to look at. The smallest interval you pick from all the tools in your feature set will determine the prediction period.

2. Advanced Options: Refining Your Prediction

Continual Learning (Beta Feature): Think of this as having a bookmark in your book. With continual learning, you can pick up right where you left off last time, building on what you've already learned. If you don't use this, you start reading the book from the beginning each time you train your model.

Weight Factor: This is like deciding how much to focus on the latest chapters of your book. A setting of 0 means you treat all chapters as equally important. Adjusting this can make your model pay more attention to recent 'chapters' of market data.

ML model: Here you choose the type of 'reading strategy' for studying your book. "XGBoost" is like a well-rounded approach that's popular for its efficiency and accuracy in making predictions.

3. Advanced Options: Training Your Model

Learning Rate: This sets the pace of your study session. A lower learning rate is like taking your time to understand each page, while a higher rate is skimming through the pages more quickly. You'll need to find a pace that's just right—not too fast and not too slow.

Total Estimators: Imagine each estimator as a group study session before a big test. This setting decides how many study sessions you have. More sessions can make you 'smarter' (more accurate), but also mean more work.

Gamma: This feature acts like a rule that stops you from over-studying to the point where you memorize the book but can't apply the knowledge. It keeps your learning effective and relevant.

Maximum Depth: Deciding the maximum depth is like setting how deep your knowledge about a subject should be. Too deep, and you might get caught up in unnecessary details. Not deep enough, and you may miss out on important points.

Minimum Child Weight: This is like setting a minimum study group size to ensure discussions are meaningful. If the group is too small, you might not get enough perspectives; too large, your study might not be as focused.

With these settings, you're adjusting how your model 'reads' the data and 'learns' from it to make the best predictions possible. Just like finding the right study techniques, fine-tuning these parameters will help your model forecast more accurately.

4. Defining Training Parameters

Train Period: This tells your model how much 'reading material' (historical data) it has to learn from. If you choose "30 days," you're giving it a month's worth of 'books' (data) to study.

Retrain: This is like periodically updating your book summary with the latest chapters. Retraining means you refresh the model's 'knowledge' with new information that has come to light since the last time it learned.

Inference Days: This is like setting a due date for when you'll need to use your knowledge. If you set it to "7," you're asking the model to make predictions about what will happen in the library for the next week based on its current 'studies' (training).

By defining these parameters, you're essentially setting up a study plan for your model, telling it how much to study, when to update its knowledge, and how far ahead it should be looking when making predictions.

5. Utilizing Feature Sets

Think of the feature set as all the different types of information and clues within a book that can help predict the story's outcome. These 'clues' include:

• Historical Price Movements: This is like looking at the plot's twists and turns in past chapters to guess what comes next.

• Trading Volume: Consider this the popularity of the book at different times—knowing when it was most read can help you understand the story's impact.

• Technical Indicators: These are like critical analyses of the book. For example, a 'moving average' is like averaging critics' reviews over time, and the 'relative strength index' could be likened to checking the intensity of the story's highs and lows.

• Macroeconomic Data: This is broader context information, similar to understanding the era or conditions when the book was written, which could affect its narrative.

These factors are the 'study materials' for your machine learning model. Just like a student uses books, notes, and reference materials to prepare for an exam, the model uses these features to learn and make educated guesses about future price movements of assets.

Feature addition: Just as you might select reference books that offer the most insight for an essay, "+ Add Feature Set" allows you to tailor the information your model studies. You can mix and match data 'chapters' to give your model the best 'knowledge' to predict the market.

Tool: Choosing your tools is like selecting the right type of reference material. Technical indicators are akin to specialized encyclopedias that dive deep into market behavior patterns, while Klines are detailed chronicles, recording every twist and turn in the market's story.

Source:

• Klines: Picking Klines is like choosing an illustrated history book, with pictures (charts) and stories (data) that show you exactly what happened at every significant moment. OHLCV—standing for Open, High, Low, Close, and Volume—provides a comprehensive picture, from the opening 'scene' of the market to its closing 'act'.

• Another Indicator: Opting for another indicator as a source is like referencing a book summary to understand broader themes. For instance, a summary of economic trends (RSI values) can inform you about the potential 'climate' of the market's future.

Evaluation Method:

Rolling Window: The Rolling Window approach ensures your model's predictions stay fresh, like regularly revisiting your study notes to make sure you remember the most current information before an exam. This dynamic 'review session' with the market data keeps your predictions in tune with the latest trends.

Timeframe Setting: Choosing a timeframe like "5m" provides snapshots of market activity in 5-minute intervals, offering a granular view. This is akin to reading a book in short, frequent sessions to catch every detail, but beware—too much detail can sometimes distract from the main storyline.

Lookback Period: Think of this as selecting the timeframe of past events you want to review for clues. Just like focusing on a specific era when studying history helps you understand how past events influence the present. The Lookback Period, is based on the smallest timeframe in your feature set so if you've selected a "5m" timeframe and a Lookback Period of "5," your model will analyze market data for the past five, 5-minute candles.

Lookback Step Size: This is akin to choosing how often you would take notes during your historical review. If you're looking over a month of market activity, deciding on a Lookback Step Size is like choosing whether to make daily summaries or only note the significant weekly events. A smaller step size means you'll have a detailed, day-to-day account of price changes, while a larger step size could give you a broader, less detailed overview.

Window Length: This is how many 'pages' of your 'book' you want to read at once. A longer window reads more pages for a bigger picture, while a shorter window focuses on the latest pages for up-to-date information.

Evaluation type: Here, you can tell the model different ways to 'read' the data, like looking at the 'differences' between data to spot changes.

Transformation Choice: Transformation is the model's way of simplifying complex market data into a format it can easily 'understand' and use. For example, a 'Linear' transformation could be likened to summarizing a dense, academic paper into a straightforward, easy-to-follow narrative.

By carefully choosing your feature sets and how the model evaluates them, you're essentially providing a well-rounded curriculum for your model to learn from, ensuring it makes the most accurate predictions based on the 'subjects' it has studied.

6. Fine-Tuning Prediction Parameters

Evaluation Metric: This is like setting a goal for what you want to learn from a book. If you choose "Percent Change," you're telling the model to focus on understanding how much a cryptocurrency's value changes over time, in percentage terms.

Lookahead Candles: Deciding on lookahead candles is like choosing how many chapters ahead in your book you want to peek at to guess what comes next. This helps the model figure out how far ahead it should try to predict the market trends based on its current knowledge.

By setting these prediction parameters, you're guiding the model on what to predict and how far ahead to look, much like planning what to study for a future exam based on past lessons.