rotaugenlaubfrosch/b3_dol_wdo

baseline subfolder:

• very simple first model to get a feeling for the data
• LightGBM tree model
• one row = one trade event
• X = pre-trade orderbook state + simple time/orderbook features
• y = next mid price - pre-trade mid price

analysis subfolder:

• goal: compare spreads between DOL and WDO over time and classify into constellations
• for every DOL orderbook row, the previous WDO orderbook row is attached
• alignment via sending_time_ns (because DOL and WDO are separate feeds)
• then the best bid/ask are compared and classified
• Conclusion:

both_1_tick_same:                      16,558,104 rows = 49.83%
dol_2_tick_wdo_1_tick_same_ask:         6,401,685 rows = 19.27%
dol_2_tick_wdo_1_tick_same_bid:         6,153,634 rows = 18.52%
both_1_tick_wdo_above_dol:                966,501 rows =  2.91%
both_1_tick_wdo_below_dol:                916,723 rows =  2.76%
both_2_tick_same:                         741,290 rows =  2.23%
dol_1_tick_wdo_2_tick_same_bid:           341,832 rows =  1.03%
dol_1_tick_wdo_2_tick_same_ask:           332,731 rows =  1.00%
dol_3_tick_wdo_1_tick_centered:            86,202 rows =  0.26%
other:                                    728,024 rows =  2.19%

Notes:

• first understand subject matter
• define X and y
• first build simple baseline model to get familiar with task
• .......
• analyze feature attribution using SHAP. Maybe new features to be derived from existing ones?
• for optimization find hyperparameters using hyperband strategy (idea)
• think about how to identify and/or mitigate concept drift and data drift after deployment
• inference latency not of concern because models can be optimized and moved to FPGA at later stage

Setup:

• OpenCode (model-agnostic & opensource) with GPT-5.5 (coding and planning), GPT-5.4-mini (for questions and explanations) and GLM-5.1 (if token limit reached)