Supply-Demand Dynamics and Regulation Techniques in Didi’s Ride-Hailing Platform

1. What is supply and demand in a transaction market?

Transaction markets consist of buyers (demand side) and sellers (supply side). The relationship between supply and demand is dynamic: when demand rises faster than supply, a shortage occurs; when supply exceeds demand, a surplus appears. Understanding these dynamics is essential for economic analysis and forecasting.

2. Didi’s business scenario and supply‑demand regulation technologies

2.1 Supply perception and prediction

Accurate perception and forecasting of supply and demand are the foundation of market regulation. Challenges include incomplete data, seasonality, sudden events, data quality issues, model complexity, and time‑lag effects.

Typical time‑series forecasting methods used are:

Smoothing methods (moving average, exponential smoothing)

ARIMA models

LSTM neural networks

Deep learning models (e.g., Transformer‑based TFT, NSTransformers, DeepAR)

Recent research (e.g., STHAN) builds a spatio‑temporal heterogeneous graph and hierarchical attention to capture complex regional relationships, achieving significant performance gains in Didi’s demand forecasting.

2.2 Integer programming for driver repositioning

The driver‑repositioning problem can be formulated as an integer programming optimization: given current driver locations, select relocation routes that minimize total cost (e.g., travel distance) while improving market balance. The workflow includes:

Candidate repositioning tasks – select drivers, target zones, expiration time, and compensation.

Task scoring – compute marginal gain of adding an idle driver to a target spatio‑temporal state.

Planning and solving – maximize total benefit under driver‑experience constraints.

Figures in the original article illustrate candidate tasks, scoring heatmaps, and driver‑gap maps.

2.3 Imitation learning

Imitation learning (IL) trains a policy to mimic expert driver behavior using supervised learning on state‑action pairs. It does not require environment interaction, making it suitable for learning from historical driver trajectories.

2.4 Offline reinforcement learning

Offline RL trains policies from pre‑collected datasets without online interaction. In Didi’s context, a driver is treated as an agent; states include local supply‑demand conditions, actions are repositioning decisions, and rewards are cumulative earnings. Methods such as AWAC, TD3+BC, CQL, IQL, and evaluation techniques like FQE are applicable.

3. Summary

Supply and demand are the core forces of transaction markets. Their dynamic interaction requires precise prediction and adjustment to maintain market balance. Didi leverages a combination of data analysis, machine‑learning models (time‑series forecasting, imitation learning, offline RL) and operations research (integer programming) to optimize driver allocation, improve passenger experience, and enhance overall platform efficiency.