IMC Prosperity is an algorithmic trading competition hosted by IMC Trading. The competition has two parts: algorithmic trading and manual trading. In the algorithmic trading section, new products are introduced each round, allowing us to implement various trading strategies. In the manual trading section, simple math and game theory questions are given. For more information, you can refer to the official wiki.

• Byeongguk Kang
• Minwoo Kim
• Uihyung Lee

The first two tradable products are introduced: STARFRUIT and AMETHYSTS. While the value of the AMETHYSTS has been stable throughout the history of the archipelago, the value of STARFRUIT has been going up and down over time.

We compared the fair price with the order book. If there was a mispricing, we execute a market order to capture the arbitrage opportunity, while making markets under normal circumstances.

Initially, we aimed to trade around the midprice. However, midprice had noise from traders placing orders at unusual prices. To mitigate this, we tried trading using micro prices (volume-weighted average midprice).

There was a slight difference between the PnL in our backtesting tool and pnl in IMC dashboard. IMC used the hidden fair value for PnL calculation. Additionally, both bid and ask sides consistently had one level with significantly large quantities. We discovered that the average of these two prices closely resembled the hidden fair value. Therefore, we traded around the average of the two prices.

Market Making:

Our goal was to place orders that maximize the expected utility per trade, calculated as (profit + other utility increment) * execution probability.

For example, if a buy order at 9998 was executed when the fair value was 10000, the profit from that trade would be 2.

Other utility includes inventory risk. When market making, keeping a position close to zero is highly advantageous. Larger positions involve higher exposure to market risk. And as positions approach their limits, the potential for profitable trades decreases due to smaller execution quantities.

Initially, we implemented Ornstein-Uhlenbeck process-based market making for AMETHYSTS and $dS=\sigma dW$ process based market making for STARFRUIT.

However, these market making models were not perfectly suited to our market since the execution probability was modeled using an exponential function. To address this, we attempted to estimate the execution probability using a Poisson distribution. Unfortunately, this did not result in a significant improvement in PnL, so we reverted back to the original model.

Market Taking:

When there were bid prices higher than the fair value, we executed market sell orders to profit from the mispricing. Also, we executed market sell orders even if the bid price was at the fair price and we had positive positions, to reduce our position.

We followed a similar approach when executing market buy orders.

ORCHIDS are very delicate and their value is dependent on all sorts of observable factors like hours of sun light, humidity, shipping costs, in- & export tariffs and suitable storage space.

We couldn't find strong connections between ORCHIDS and the various factors. So we didn't do directional bets or price prediction.

Initially, we considered market making due to the wide bid-ask spread of ORCHIDS, but found it impossible because bids lower than the fair price and asks higher than the fair price did not fill well.

Instead, ORCHIDS were also tradable on the exchange of the south archipelago. Therefore, we implemented arbitrage between the two exchanges. We placed orders that maximizes expected profit per trade (= (enter price on this island - exit price on the south archipelago - shipping cost - import tariff) * execution probability). We estimated the execution probability through our own experiment.

The GIFT_BASKET is now available as a tradable good. This lovely basket contains three things:

1. Four CHOCOLATE bars
2. Six STRAWBERRIES
3. A single ROSES

All four of the above products can now also be traded on the island exchange. Are the gift baskets a bit expensive for your taste?

The price of GIFT_BASKET oscillated based on the underlying index, so we only traded GIFT_BASKET. We did not trade the underlying assets to reduce transaction costs. Although trading the underlying assets could have made the portfolio more stable despite some transaction costs, our goal was to maximize expected returns, so we chose not to do so.

We optimized the trading strategy by augmenting data using Monte Carlo simulation. This helps us avoid overfitting to past data.

When trading GIFT_BASKET, we placed market orders deep into the order book, accepting slippage to seize trading opportunities immediately.

Since our limit orders were rarely filled, we did not performed market making.