Chapter 178: Asynchronous Federated Learning for Trading

Overview

In the previous chapters, we used synchronous aggregation (FedAvg). While robust, synchronous FL has a fatal flaw in low-latency trading: it waits for the slowest participant (the “straggler”).

Asynchronous Federated Learning (AFL) allows the central server to update the global model as soon as any single client’s update arrives.

The Challenge: Model Staleness

In AFL, a client might start training on version $N$ of the global model, but by the time their update arrives, the server has already moved to version $N+10$. This is called Staleness. If we treat outdated updates the same as fresh ones, the model will diverge.

Our Solution: Staleness-Aware Aggregation

We will implement an aggregation rule that automatically down-weights “stale” updates using a decay function:

Fresh updates: High influence on the global model.
Outdated updates: Low influence, used only to preserve long-term patterns.

Project Structure

178_asynchronous_fl_trading/
├── README.md           # English Overview
├── README.ru.md        # Russian Overview
├── docs/ru/theory.md   # Mathematical deep-dive
├── python/
│   ├── model.py            # Base Neural Network
│   ├── afl_core.py         # Asynchronous server logic
│   └── train.py            # AFL vs. Sync FL simulation
└── rust/src/
    └── lib.rs              # Optimized decay coefficient engine