Cashew Capital Fund - Automated Trades

Traditional trading methods can be time-intensive, struggle to adapt to rapidly shifting market conditions, and often result in inefficiencies.

🥜 Cashew Capital Fund is a startup investment management fund interested in using ML to address this gap.

Our Objective: Design a machine learning algorithm that identifies trading opportunities through technical analysis and autonomously determine the best times to enter and exit for any particular stock.

Data Preparation

Data Source: Benzinga Time-Series Data

• Historical (5-min candles): Benzinga GET Bars
• Real-Time (15-min delay): Benzinga Delayed API

Technical Indictator	Formula	Fine-tuned Parameters	Explanation
Exponential Moving Averages (EMA)		N = 36-period & 78-period (3 and 6.5 hours)	Stabilizes trends for 2-day predictions Better than standard spans of 12 & 26 days
Relative Strength Index (RSI)		9-period (45 min)	Optimized for voilatile markets
Moving Average Convergence Divergence (MACD)		27-period (Signal Line)	Focuses on sustained signals, aligning with longer EMAs
Stochastic Oscillator		9-period (SlowK) 3-period (SlowD)	(SlowK) Increases sensitivity to recent price action (SlowD) Balances responsiveness with noise reduction

Algorithmic Approach

Features: MACD, MACD-Hist, RSI, SlowK, SlowD

Note

We choose a classification approach. Why?
Trading decisions are categorical: you either buy, sell, or hold. Since we want an approach that is (1) easy to interpret and (2) quick to implement (for quicker testing and given limited timeframe), classification was the way to go.

Perfect for high frequency trading! ✨

---

Method: Create a label called “signal” that represents whether we should buy or sell a stock based on its price over the next 2 days

Signal Type	Conditions
Strong Sell (-2)	Min price ≤ 0.95 CP Maximum price from now until the Min position ≤ 1.01 CP
Weak Sell (-1)	Min price ≤ 0.97 CP Maximum price from now until the Min position ≤ 1.01 CP
Strong Buy (2)	Max price ≥ 1.05 CP Minimum price from now until the Max position ≥ 0.99 CP
Weak Buy(1)	Max price ≥ 1.03 CP Minimum price from now until the Max position ≥ 0.99 CP
Hold (0)	Not Applicable

Legend

• CP = Current Price
• Min = Min price over 2 days
• Max = Max price over 2 days
• Threshold = 5%
• Future_period = 155
• Threshold_weak = 3%
• Min_diff = 1%

---

Here’s a glimpse of our algorithm applied to approximately 180 days of NVDA data:

Model Selection

Model	Reasoning
Random Forest Classifer (RF)	• Less prone to overfitting due to using multiple decision trees instead of one • Resistant to noise and outliers
Gradient Boosting Classifer (GB)	• Tends to achieve higher accuracy compared to Random Forest • Excellant capturing complex relationships between features and the target variable

Hyperparameter Tuning

Warning

There was an error in how we configured and ran RandomizedSearchCV, which affected the hyperparameter sampling process. As a result, the reported results may be skewed and should not be considered fully reliable. We recommend rerunning the search with corrected parameters to ensure accuracy.

Tuning Method	Accuracy Score
RF - None	0.67
RF - GridSearchCV	0.54
RF - RandomizedSearchCV	0.54
GB - None	0.55
GB - GridSearchCV	0.60
GB - RandomizedSearchCV	0.59

NoSQL Database

Since our dataset updates daily, we used MongoDB to store our data on the cloud (better accessibility ✅) The database was organized into 3 categorical collections:

├── Historical Data/
├── EMA-78 Data/
│     └── AAPL-78
│     └── MSFT-78
│     └── TSLA-78
│     └── ...
├── Market Logs/

• Historical Data: Stores historical stock data for the last 6 months, with 5-minute candlestick intervals (this is what the model is trained on)

• EMA-78 Data: Stores 78 rows of past data needed to compute the technical indictator during real-time excutation

• Market Logs: Stores logs of model executation (buy or sell). The schema is as follows:

  • Timestamp
  • Trade Type
  • Monetary Metrics (e.g., trader’s funds and stock prices)
  • Volume

Installing Dependencies

To install the necessary dependencies for this project, follow these steps:

1. Install the dependencies:

   pip install -r requirements.txt

2. Verify installation:

   pip list

Credits and Acknowledgements

Special thanks to Swathi Senthil, George Abu Daoud, Bharath Venkataraman, and Boshen Parthasarathy for the mentorship and feedback.

Tools/Libraries: Google Colab, VSCode, Scikit-learn, MongoDB, Benzinga, Slack