How to Use

Running Neural Network Model

Step 1: Open the Installed Application

Step 2: Navigate to "Run Neural Data Update"

Step 3: Click "Run Neural Data Update" to run the model and make inferences

Information: This step may take several minutes to run

Step 4: Neural Network Inference Analysis

🤖 Neural Data Update (DFT Output)

📊 What This Section Is

This module runs the Deep Factor Transformer (DFT) neural network model and outputs correlation scores between each stock and the model’s learned factor signals.

🔍 What Information It Shows

🧠 How to Interpret the DFT Correlation

In this case:

• SMCI has the highest score (0.1418), suggesting the DFT model finds this stock’s features profitable.

🎯 Purpose of This Module

• Ranks stocks by signal confidence from the neural model.

• Feeds directly into the backtesting module, where the top-K stocks based on these scores are selected (and N randomly dropped).

• Helps diagnose the quality and breadth of the model’s signal.

Running Backtesting System

Step 1: Open up the Installed Application and Navigate to "Backtest"

One can set these parameters as they wish to. The model cannot go past March 31'st 2020, and it cannot go any later than the current date.

Top K and Drop N are parameters for Qlib's Strategy

Here is an overview:

This strategy selects the top K instruments based on model scores and then randomly drops N of them before actually trading.

Top K

• The number of instruments to select per trading day with the highest model score.

• Example: If Top k = 10, you look at all signal scores on a given trading day and pick the top 10 instruments to consider for trading.

Drop N

• From the top K selected instruments, randomly drop Drop N before constructing the final portfolio.

• This adds regularization by encouraging diversification and reducing overfitting to model scores.

• Example: If Drop N = 2, and Top K = 10, only 8 instruments will be used to form the portfolio on that day (randomly chosen from the top 10).

Once all this is selected press Run Backtest. This step may take up to several minutes.

Step 2: Analyzing Backtest Results

📊 Sections Explained

• bench: Performance of the benchmark (e.g., SP500).

• no_cost: The model's performance ignoring trading costs

• with_cost: The model's performance after including trading costs (Transaction Costs.. etc).

📈 Metrics Explained

📊 What This Graph Is

This is an equity curve — a plot of cumulative returns over time for the model after transaction costs (as selected in the dropdown: "(With Cost)"). One can select "No Cost", and also "Benchmark" to see their respective charts

🔍 What Information It shows

📈 Equity Curve: Model (With Cost)

✅ Why This Is Important

• Trend direction: Upward = profitable model.

• Volatility: Sharp rises or falls reflect risk and potential drawdowns.

• Drawdown depth: Tells you the worst-case dips from peak (important for risk tolerance).

• model stability: Smooth curves suggest consistent performance, while jagged ones may indicate overfitting or market sensitivity.

📈 Equity Curve: Model (No Cost)

✅ Why It Matters

• Shows the maximum possible performance of the model.

• Serves as an upper bound for expected returns.

• The difference between No Cost and With Cost curves helps quantify how much trading costs eat into profits.

📉 Equity Curve: Benchmark

✅ Why It Matters

• Provides a baseline for comparison.

• Helps determine whether your model is adding alpha (i.e., outperforming the market).

• If the model consistently stays above this line, it's a sign of value-add.

📉 Trade Return Distribution

📊 What This Graph Is

This is a histogram showing the distribution of individual trade returns from the model (specifically With Cost). It tells you how frequently different return sizes occurred across all trades

📋 Trade Summary Panel

📊 What This Section Is

This panel provides a quick statistical overview of the model's individual trade-level performance over the backtest period.

🔍 What Information It Shows