RNN versus Transformer policies for a small trading experiment.

01 · Experiment

Can a policy learn when to buy or sell from a price sequence?

The project framed trading as an OpenAI Gym environment. The state described the market observation; actions changed the portfolio; reward reflected the resulting gain or loss. Two deep-learning policies—an RNN family model and a Transformer—were compared under the same educational setup.

02 · Policies

Sequential memory versus parallel attention.

The report intentionally omitted a Transformer decoder: the policy needed a representation for action selection, not a translation back into another sequence.

03 · Reported results

The Transformer led in this run—but the numbers require context.

Values above are transcribed from the attached report. They are historical experiment outputs, not independently rerun results and not evidence of future returns.

04 · Reading the results

The Transformer won this experiment, not the broader argument.

In the chosen setup, the Transformer policy earned the higher reported profit and reached the target portfolio ratio in fewer episodes. Using a held-out time period was also better than reporting training performance alone. The experiment therefore shows that attention was a credible alternative to recurrent processing for this task.

It does not show that Transformers are better trading systems. The comparison used one index, one historical window, 10 training episodes, and no repeated random seeds. It did not account for transaction costs, slippage, drawdown, or risk-adjusted return, and stopping evaluation at a profit threshold may have favored the apparent winner. A defensible conclusion is narrower: the Transformer outperformed the RNN under this environment and stopping rule; whether that result generalizes remains unknown.