This repository is Zhi Yun Yap's solution to the programming project for Mid-To-Long-Term Quantitative Researcher.

conda create -n project python=3.7
conda activate project
pip install -r requirements.txt

Only standard libraries are used in this project.

To generate position assignment, run,

cd src
python3 run_assignment.py --file <FILENAME> --output_dir <DIR_PATH>

This generates position assignment and saves the results to FILENAME under output_dir.

To replicate results, run

python3 run_assignment.py --file target_position.csv --output_dir results

To evaluate position assignment, run

cd src
python3 run_evaluation.py --file <FILEPATH> --output_dir <DIR_PATH>

This prints a summary of the portfolio performance and saves corresponding charts under output_dir.

To replicate results, run

python3 run_evaluation.py --file target_position.csv --output_dir results

• Load and reformat data from zip file
• Construct reversal signals at every minute
  • Bollinger band
  • Stochastic oscillator
  • MACD
  • Momentum
• Fill missing price with last available price

Taking momentum signal as an example, at the end of each minute, we rank stocks from lowest to highest trailing return. We then split the top/bottom half names into long/short bucket. For each stock in the long/short bucket, we can assign either (1) equal weighting, or (2) signal weighting to obtain the weights.

1. Equal weighted

   Each stock in the long/short bucket is assigned equal proportion of the notional value

2. Signal weighted

   Normalized signal scores within long/short bucket by sum of signal scores in each bucket

We start by constructing univariate model - use one reversal signal to weight position and evaluate the efficacy of individual signal to assigninig target position.

The final position assignment model is a multivariate model -- take the average of Bollinger band and Momentum values as the new signal, and assign position using a signal-weighting scheme.

The model is developed with the following properties:

• Zero net exposure (measured in notional value) minute-by-minute
  • By normalizing signal scores within long/short bucket
• Zero overnight risk exposure
  • By liquidating all positions by market close

When constructing and evaluating models, we made following assumptions:

• Frictionless market - ignore transaction cost, bid-ask spread, commission fee
• Able to rebalance portfolio and adjust position by the end of each minute interval
• Able to enter a position at open price in next minute

We evaluate overall performance of the position assignment model by evaluating the corresponding portfolio using

• Sharpe ratio
• Annualized volatility
• Max drawdown
• Max daily turnover
• Average number of assignments

We also inspect how portfolio PnL (overall, long/short bucket), return, volatility, turnover, and wealth curve evolves over time.

The repository is structured as follows:

• Main modules
  • data
    • dataloader: handler to load and transform data
  • models
    • base: abstract class for position allocator
    • position_allocator: concrete class for position assignment model
    • backtest: evaluate portfolio performance
  • results: store position assignment and evaluation results
• Others
  • utils: helper functions
  • signals: compute technical indicators as reversal signals

• Evaluate sensitivity of portfolio performance with trading friction
  • The frictionless market is an ideal condition and can be unrealistic
  • Investigate how performance deterioriate under varying transaction cost from (e.g., 0bps - 30bps)
• Train a ML model (e.g. XGBoost) to select and combine technical indicators to generate more optimal position assignment