Why Neural Networks Don't Predict the Moscow Exchange, and Where ML Actually Helps Traders

At a machine learning in trading meetup in Moscow, one sober thesis stood out: neural networks are almost useless as a tool for direct price forecasting. ML does, however, help in narrower tasks: identifying the market regime, filtering out bad trades and cutting execution losses.

The illusion of price forecasting

A popular fantasy still circulates in the market: show a model the chart of Sberbank or an index, and it will output the exact price for tomorrow. In practice, this falls apart very quickly. Price is too noisy, the market keeps changing, and any pattern found in historical data often vanishes the moment the strategy goes live with real money. In high-frequency trading the situation is even harsher: nanoseconds matter, and a heavy model can simply kill the system with latency.

"ML does not give you control over the market and is not a magic money button."

That is why experienced market participants see machine learning not as a generator of ready-made signals, but as a supporting layer on top of a base strategy. Even if a model guesses the direction of a move more often than chance, that is not enough. Between a good forecast and actual money sit the order book queue, slippage, and broker and exchange fees. As a result, a formally strong model can lose money where the backtest looked almost perfect.

Where ML works

The soundest idea from the discussion is simple: machine learning pays off where the task is narrow and verifiable. Instead of trying to predict the market itself, the model helps break it into components: which regime it is in, which signals to skip, and where costs will eat the potential profit. This is much closer to the real practice of a solo algo trader, who cannot afford a hedge fund's infrastructure.

• Market phase detection: trend, sideways range, volatility spike
• Signal filtering: which trades from the base strategy are better skipped
• Execution optimization: how to enter a position with less slippage
• Order book microstructure analysis: finding patterns in the order flow

This matters especially for a retail trader, who often plays every role at once: sourcing data, cleaning it, designing features, building the model, testing hypotheses, calculating risk, and then pushing the result into production alone. In that setup, ML is useful not as a showcase of big promises but as a tool for a specific, narrow improvement, such as switching off a trend strategy in a hard sideways market or skipping a trade with poor expected execution.

Why strategies die

Most problems start not in the model but earlier, at the data and hypothesis-testing stage. If a day is missing from the candles, futures are stitched together incorrectly, or features accidentally contain information from the future, the algorithm instantly finds a pattern that does not exist. On historical data it looks like the holy grail, but in live trading it ends very quickly. Hence the main principle: garbage in, garbage out. Data quality and an honest experiment matter more here than the choice between fashionable libraries and architectures.

The discussion yielded a useful five-question filter to run before writing any code: is there a concrete task, is there honest data, can the idea be tested with fees and latency included, does the forecast turn into a real action, and does all of this produce an economic effect. If even one question lacks a confident answer, ML is being brought in too early. Otherwise you fall into the typical trap: there seems to be a forecast, but after all costs there is no profit.

What this means

For the market, this is a healthy sign of sobriety. Neural networks do not replace a trading system and do not eliminate risk, but they help improve processes at specific points where a human is already hitting limits on data volume or reaction speed. Another important conclusion is that the market lacks experience sharing: while solo traders repeat the same mistakes separately, ML remains not an accelerator of profit but an accelerator of expensive experiments.