Future · 9 min
When most of the order flow is bots: what actually changes
The interesting question is not whether AI-driven bots take over crypto markets. It is what happens to your edge when everyone else has one too.
Engineering blog
50 pieces on how trading bots actually work and where automated trading is heading: Bybit V5 API, WebSocket, risk engines, honest strategy validation, AI in trading, and the decisions that come before the code.
We write about what we have built with our own hands. There are no 'top 10 strategies' listicles here.
Featured
The interesting question is not whether AI-driven bots take over crypto markets. It is what happens to your edge when everyone else has one too.
Not “ask a model what to buy”. A working ML strategy is a data pipeline, a labelling decision, honest validation and an expected-value gate — the model is the smallest part.
A marketplace bot with a beautiful backtest is not lying to you exactly. It is showing you the one configuration, out of thousands tried, that happened to fit history.
Where automated trading is heading, what changes when models make the decisions, and which parts of that are already real.
The interesting question is not whether AI-driven bots take over crypto markets. It is what happens to your edge when everyone else has one too.
Not “ask a model what to buy”. A working ML strategy is a data pipeline, a labelling decision, honest validation and an expected-value gate — the model is the smallest part.
When most participants run similar models on similar data, they stop being independent actors. That correlation is invisible in normal conditions and decisive in bad ones.
An LLM is excellent at turning unstructured text into structure and terrible at being the thing that decides to send an order. Those are different jobs.
Every edge has a half-life. What changed is that the tooling to find and copy an edge got cheap, so the half-life got shorter — and planning for that is part of the design.
Deep learning dominates domains with enormous data and high signal-to-noise. Market prediction is the opposite of both, which is why boosted trees on hand-built features keep winning.
Strategies rarely fail loudly. They degrade, and the equity curve is the last place it becomes obvious. Detecting it earlier is a monitoring problem with a known solution.
RL is the most requested and least delivered approach in retail quant. The blocker is not the algorithm — it is that you cannot build the simulator it needs.
Everyone wants AI to decide what to buy. The higher-value and lower-risk application is having it decide when to stop.
Explainability in trading is not an ethics requirement. It is what determines whether a human can make a correct decision about the system at 4am.
You cannot generate more history. What you can generate is plausible alternative history — which is excellent for finding fragility and terrible for finding edge.
Averaging models reduces variance if the errors are uncorrelated. On market data they usually are not, which is why most trading ensembles are one model with extra latency.
The decisions before the code: whether to automate at all, build or buy, what a bot should cost and how to specify one.
A marketplace bot with a beautiful backtest is not lying to you exactly. It is showing you the one configuration, out of thousands tried, that happened to fit history.
Most clients never open the repository. It still changes the relationship, because everything that can go wrong with a vendor is survivable when you hold the code.
Every arrangement where someone else's infrastructure trades your account concentrates risk in a place you cannot see. There is no technical reason to accept it.
Small size limits how much a bug costs. It does not make the bug appear on schedule, and the bugs that matter only show up in situations you cannot arrange on demand.
We guarantee that the software behaves as specified. We do not guarantee what the market does with it. Anyone collapsing those two is selling something other than software.
The comparison people make is a subscription against a development quote. The comparison that matters includes what you can change, what you own, and what happens when the strategy needs to move.
We turn down or rescope a meaningful share of enquiries. Here is the list of conditions that make automation the wrong next step, written so you can check yourself before paying anyone.
The same sentence — “a Bybit bot” — covers a week of work and six months of it. Here is what moves the estimate, so you can shape the scope instead of receiving a number.
They are the same six every time, they are all cheap to avoid, and each one is usually discovered the expensive way.
A good spec is not long. It answers a fixed set of questions precisely, and writing it will improve your strategy before anyone writes a line of code.
V5 REST and WebSocket in practice: signing, streams, fills, limits and the errors that survive testing.
One account model, one set of endpoints, and a handful of details that silently break bots ported from earlier versions.
A dropped WebSocket is not an error to log. It is a moment when your bot's picture of the world and reality quietly diverge.
A limit order that fills halfway leaves you with a position nobody planned, a stop sized for a position you do not have, and an average price your code has not recomputed.
Limits never bite when the market is calm. They bite in the minute when your bot urgently needs to cancel, and that is the minute the design has to be built for.
Two independent legs on one symbol sound like more control. Sometimes they are. Often they are a way to pay funding on both sides while feeling hedged.
Most key incidents are not sophisticated. A key ends up in a repository, a chat message, or an environment that more people can read than anyone intended.
How a production trading system is put together: layers, order state, risk, sizing and monitoring.
Almost every bot that becomes unmaintainable made the same mistake — it let strategy logic reach directly for the exchange.
The exchange telling you it received your order and the exchange telling you the order traded are different events. Bots that conflate them report profit that never existed.
Risk management in a trading bot is not a calculation. It is a set of refusals, enforced in code, that the strategy cannot argue with.
The single most useful habit in automated trading is computing size from the distance to your stop. It is also the one most bots get almost right and then lose to fees.
P&L tells you what happened. It does not tell you whether the machine that produced it is working — and by the time P&L reveals a broken bot, it has been broken for a while.
Multi-venue support is not about supporting more venues. It is about having exactly one place in the codebase that knows how any particular venue is weird.
Recording everything is what makes a system honest. Recording it synchronously is what makes it miss the market. The fix is boring and worth getting right early.
Proving that an edge exists before risking money on it — honestly.
A backtest tests a hypothesis about the market. Paper trading tests your software against reality. Skipping either one means shipping on an untested assumption.
Every leaked bit of future information makes a backtest look better. That is why leaks are so hard to notice — the evidence of the bug is a result you were hoping for.
Latency only matters in proportion to how fast your edge decays. Most strategies are not latency-sensitive and spend money as though they were.
Standard gradient boosting treats rows as exchangeable. Market data is not. We wrote a library where purged validation is the default and look-ahead is a lint error.
Assuming you fill at the touch is the most common way a backtest lies. A crude but measured slippage model is worth more than a sophisticated guess.
Splitting time-ordered data at random produces beautiful, meaningless results. The correct procedure is not much harder and it is the difference between a result and a story.
How perpetuals, liquidity and order flow actually behave, and what that means for a bot.
A perpetual has no expiry, so something has to keep it tethered to spot. That something is a payment between longs and shorts, and it is a real line in your P&L.
A market maker earns the spread and pays for information. Every fill you get is a fill somebody wanted, and the whole discipline is about surviving that asymmetry.
The book contains far more numbers than signal. A handful of constructions carry most of what is usable, and most of the rest is a slower way to look at price.
A cascade is not a market opinion. It is forced selling by an exchange algorithm, and it is the reason a stop-loss can fill several percent past its level.
Open interest tells you how much leverage is in the market. It is genuinely informative and it is also the single most common source of look-ahead bias we find.
Both have a high win rate and a payoff shape that hides the loss in the tail. That is not a reason to avoid them — it is a reason to build them with a hard ceiling.
Opinion, argued from the same engineering experience as everything else here.
Not because a bot is smarter than you. Because it is the same on Tuesday at three in the morning as it was on Monday at noon, and you are not.
A bot has no intuition, no memory of yesterday's fear, and no ability to notice something you did not tell it to look at. Understanding that is how you write a good one.
Markets are the one place where a large number of automated agents already compete against each other with real money. What has happened there is worth more than most predictions.
We write about what we build. If you need it built, get in touch — scoping is free.