A3T-GCN Gets Better FTSE100 Forecasts from Shorter Histories

If you buy or sell stocks, even a small edge in tomorrow’s price forecast can matter. This paper tests A3T-GCN, a hybrid model that combines attention with a graph convolutional network (a neural network that learns from relationships between companies) to predict closing prices for FTSE100 constituents. The dataset covers 79 companies and 375,329 daily observations from 2007 to 2024, using features such as RSI, MACD, normalized returns, log returns, and annualized log returns over 1 week, 2 weeks, 1 month, and 2 months. The model builds graphs from sector classifications and correlations of returns or financial ratios. The headline result is simple: annualized log returns and shorter input sequences improve prediction accuracy while also reducing computing requirements. Longer histories help, but only modestly, which suggests they matter more when the forecast horizon stretches further into the future. In other words, the model seems to get more useful by focusing on the right recent signals, not just by looking farther back.

375,329 daily observations sit behind this forecast test. They cover 79 member companies in the FTSE100 index from 2007 to 2024. That is a huge diary of market moves. The surprise is not that more history helps. The surprise is that shorter slices of history helped more. The model used the right kind of return signal. If you track stocks, that matters. It suggests the freshest clues can beat a longer memory when the goal is tomorrow's close. A model like this does not read prices like a line on a chart. It also reads links between companies. A graph turns neighbors into clues. That is where graph learning earns its keep.

When less history helped more

The A3T-GCN model did best with annualized log returns. Those are returns scaled to a full year. They give short windows a common yardstick. The shorter input lengths also won. They improved accuracy and cut computing power at the same time. Longer histories still helped, but only a little. The gain from longer context was modest. That pushes against the idea that more past data always gives a big boost. Across these tests, the cleanest signal came from the recent past. That trade-off is the heart of the result. A faster model is easier to test again and again. It also uses less compute inside a larger pipeline.

How the market graph was wired

A3T-GCN blends attention with a graph convolutional network. Attention is a way for a model to focus on the most useful parts of the input. A graph convolutional network is a model that learns from links, not just rows in a table. Here, the links came from sector labels and from return or ratio correlations. Pearson correlation checks straight-line movement. Spearman correlation checks shared rank order. RSI, or relative strength index, tracks momentum. MACD, or moving average convergence divergence, tracks trend shifts. The node features also used normalized returns, log returns, and annualized log returns. The model then read each company in context, not in isolation. That matters when one name moves with its peers.

Why shorter windows matter

For anyone trying to forecast stocks, the model points to a practical trade-off. You do not always need the longest possible memory. You need the signals that stay useful inside the graph of related companies. Shorter input windows also trim compute. A forecast run costs less time and power. Teams can test the same setup across many names. The result also gives annualized log returns a strong case as a price feature. Those inputs gave the clearest lift. That makes them a better bet than raw, untidy past moves alone. It also makes the whole system easier to repeat. That matters for live work. It saves time on each run.

What to test next in FTSE100

The surprise now has a clear use. The model did better with shorter histories and annualized log returns. That means a forecast tool can get leaner without giving up accuracy. It can also save compute, which helps when the same setup runs across many FTSE100 names. The open test is more distant future predictions on the same 79-company set. The abstract says longer history matters more there. So the real question is where the short-memory edge stops paying off. That boundary is the piece model builders will want to map next.