HFT market making in pytorch and gym

Agent

Environment State Space is consists of eight indicators:

1. LOB levels rendered as stationary prices;
2. cumulative notional value at each price level;
3. imbalances for cumulative notionals;
4. order inflow imbalances at LOB levels;
5. the spread between the best bid and ask prices;
6. custom price momentum indicator;
7. custom trade flow imbalance indicator; and
8. environmental reward signal.
9. forecasted volatility.

Distance to midpoint

Distance $\xi$ is the difference from price $p$ at LOB level $i$ from midpoint $m$ at time $t$. $$ \xi_{t,i}^{bid,ask} = \frac{p_{t,i}^{bid, ask}} {m_t} - 1 $$

Cumulative Notional Value at Price Level

Cumulative sum $\chi$ is the sum of the product of price $p$ and quantity $q$ at LOB level $i$ at the time $t$ for both $bid$ and $ask$ sides. The LOB's first 15 price levels are extracted from $bid$ and $ask$ sides of the LOB, thus creating 30 values for this features. $$ \chi_{t,i}^{bid, ask} = \sum_{i=0}^{I-1} p_{bid, ask}^{t,i} \times q_{bid, ask}^{t,i}$$

Notional Imbalances

The order imbalances $\iota$ are represented by notional value and normalized to the scale [−1, 1]. Since there are 15 price levels included, there are 15 values for this feature. $$ \iota_{t,i} = \frac{\chi_{t,i}^{ask,q} - \chi_{t,i}^{bid, q}}{\chi_{t,i}^{ask,q} + \chi_{t,i}^{bid,q}} $$

Tools to use

1. pytorch
2. cryptofeed
3. influxdb
4. DQN
5. Trading Gym

How to use

Train:

python experiments.py --training true

Testing

python experiments.py --training false