Quote Craft

Project Structure
│
└─── Part 1 : BTCUSDT-26Sep25 (Bitcoin derivative contract against Tether, with expiry on 26th Sept 2025)
│   │   
│   └─── Started with a Naive Market Making Strategy, as a benchmark.
│   └─── Compared it with our Inventory Aware Market Making Strategy.
│   └─── Implemented realistic slippage and execution.
│   └─── Tested our strategy on three consecutive days. 
│
└─── Part 2 : (Further Improvements)
│   │   
│   └─── Bayesian Hyperparameter Tuning using Optuna
│   └─── (future) Paper Trading using Alpaca API

Part 1

This part of the project aims to implement an intelligent market-making strategy, generating calculated quotes on order book data. It is structured as Python notebooks designed to simulate quoting, execution, and inventory management in a high-frequency trading environment. The simulator aims to help understand and evaluate simple market-making behaviors under realistic market conditions.

1_1_Market_Making.ipynb

Contains the base version of the simulator. Quotes are placed deterministically at spread-adjusted prices, and fills are assumed to happen immediately when a price match occurs. Useful for understanding core logic: quoting, execution, inventory management, and mark-to-market PnL tracking.

• Naive Market Neutral Strategy (benchmark)

  • Quote Generation : $$min\_price : \Delta = 0.1$$
    • Price
    $$bid\_quote = best\_bid + \Delta$$ $$ask\_quote = best\_ask - \Delta$$
    • Volume
    $$base\_volume : Q_0 = 0.001$$ $$bid\_vol = Q_0$$ $$ask\_vol = Q_0$$

  (Performance Ratio = 68 %)

• Inventory Aware Smart Strategy

  • Quote Generation : $$min\_price : \Delta = 0.1$$ $$price\_factor : \alpha = 3.0$$ $$Inventory\_bias : \phi = \min(\frac{|I_{volume}|}{I_{margin}},\alpha)$$ $$I_{margin} = 10$$
    • Price
    $$bid\_quote = best\_bid + \Delta - \lfloor\phi\rfloor . \Delta$$ $$ask\_quote = best\_ask - \Delta - \lfloor\phi\rfloor . \Delta$$
    • Volume
    $$base\_volume : Q_0 = 0.001$$ $$vol\_factor : \beta = 0.25$$ $$bid\_vol = Q_0 (1 - \beta . \phi)$$ $$ask\_vol = Q_0 (1 + \beta . \phi)$$
    • NOTE :
    $$\alpha \cdot \beta \lt 1.0$$

  (Performance Ratio = 79 %)

This strategy saves us from the risk due to holding a big position.

1_2_Realistic_Market_Making.ipynb

Extends the base simulator with two realism enhancements:

• Execution Probability: Not all limit orders are assumed to be filled. Each order has a probabilistic chance of execution, simulating queue position and competition.
• Gaussian Slippage: When trades are executed, the price is adjusted with a small random noise to reflect market slippage and non-instant execution.

As expected, this leads to more volatile PnL outcomes, and emphasizes the risk of non-execution and price impact.

Part 2

Automated Hyperparameter Tuning : 2_1_HpTuning_Cold.ipynb

• We to build a personalised Cold_Tuner. The name 'cold' because it takes some time (~5-15 minutes) to find the parameters, without using any initial guess.

• It uses Optuna's TPESampler (Tree-structured Parzen Estimator).

  Details:

  • It's a Bayesian optimization method based on probabilistic modeling.
  • It builds two models: one for "good" hyperparameter values and one for "bad" ones, and then chooses values that are more likely to improve performance.
  • Works well for non-convex, noisy, and high-dimensional search spaces. (default sampler in Optuna)

• To improve the time taken, we use Optuna's Median Pruner.

  Details:

  • Since our Realistic_Market_Making is randomised. During the HpTuning, we might end up getting a set of Hps which performed well once (got lucky), but don't reflect a robust and promising strategy otherwise.
  • To prevent against this, we run MAX_REPS = 7 times and report the $\frac{mean(score)}{std(score)}$ to the tuner.
  • This would demand 7 times the compute. To reduce that factor, We prune the runs with a poor score starting from PRUNE_AFTER=3.

Hyperparameters we're interested in, and their typical ranges :

• Price_factor discrete vales in $[1,\infty)$

• Vol_factor continuous values in $(0,1)$

• inventory_cap continuous values in $(0,\infty)$

Score :

$$score = \frac{final\_pnl}{std\_dev(pnl[t] - max\_theo[t]) \cdot std\_dev(Inv\_vol)}$$

(The Metric which optuna maximizes)

Potential Improvements :

• The hidden variables of Market are bound to drift with time.

• We can build a fast Hot_Tuner to frequently adjust the parameters. (more like a re-tuner)

• One could use local search on the initial guess given by the previous iterations, with Optuna's GPS Sampler

Paper Trading on Alpaca : (Future)

• Alpaca does not allow one to have a negative position in some instruments, including Bitcoin. This would not let us do Market Making as we want to.

• Moreover, the contract of our choice (BTCUSDT-26Sep25) is not available on Alpaca. So we are chose some other instrument.

• Moving to another instrument implies having to tune the hyperparameters again. Our Hp_Tuner saves the manual effort.

• So let's chose Coca Cola (KO) because it's my favorite drink ;)