← Back to blog

Before Your First Bot

Published 2026-07-23 · By Drew Shelem

A trading bot does not make decisions. It repeats yours, quickly and without flinching.

That sentence is the whole thing, and almost every expensive mistake people make with bots comes from forgetting it. The bot is not looking for opportunities on your behalf. It is not smarter at three in the morning than you are at three in the afternoon. It executes a rule you wrote, thousands of times, exactly as written — including the parts you got wrong.

So the interesting question was never “how do I build a bot.” It is “is the rule any good,” and that question has a specific answer that most people never compute.

What a strategy actually is

Strip away the vocabulary and a strategy is four decisions:

When to enter. Some condition in the data that you believe carries information — a price move, a time of day, a relationship between two markets.

At what price and size. How much you are willing to pay, and how much of your capital goes in.

When to exit. At a target, at a loss, at a time, or by holding until the market resolves itself.

When not to trade at all. The most underrated of the four.

That is it. Everything else — the code, the hosting, the dashboard — is plumbing. Two people with the same four decisions have the same strategy regardless of what they built it with.

Where the money is supposed to come from

Not from automation. Automation is a speed and consistency advantage, and speed only pays if you are pointed in a profitable direction to begin with.

On a prediction market the price is a probability. A contract trading at 50¢ is the market saying “we think this is about 50% likely.” That price is not arbitrary — it is what everyone else believes, weighted by how much money they were willing to put behind it.

Your profit comes from one thing: your forecast being better than that price, often enough, by enough. The gap between what you think the probability is and what the price says it is — that is called your edge, and it is the only source of return in the entire system.

This is worth sitting with, because it has an uncomfortable implication. If you have no view that differs from the market’s, no amount of engineering produces a profit. A very fast bot with no edge loses money faster than a slow one.

Three costs that do not care about your opinion

The exchange fee. Every trade you take costs something. On Polymarket at a 50¢ price on a crypto market, the fee is worth about 1.75 percentage points of probability. In plain terms: to break even at 50¢ you do not need to be right 50% of the time. You need to be right about 51.75% of the time. Everything below that is a slow loss no matter how many trades you make.

The spread. The price to buy and the price to sell are not the same. The difference is a cost you pay on entry, every time, and it is frequently larger than the fee.

Being wrong. The obvious one, and the one people budget for. The first two are the ones that quietly decide the outcome.

Why a week of results tells you nothing

This is the part that costs people the most money, and it is not intuitive.

Suppose your strategy genuinely has a small edge. Over a week you might see a loss. Over a month you might see a loss. Not because the edge disappeared, but because a small edge is buried under a large amount of random variation, and it takes a great many trades before the signal separates from the noise.

The number is often shocking. A real, genuine two-percentage-point edge on markets priced near 50¢ can require on the order of a hundred thousand trades before you can say with statistical confidence that it exists at all.

This cuts both ways, and the second direction is the dangerous one. A profitable week is equally uninformative. Most people who switch off a losing bot after two weeks would have switched on a losing bot after two profitable ones. Both decisions are made on the same non-evidence.

The practical consequence: you cannot evaluate a strategy by watching it. You have to test it against history properly, and you have to decide in advance what result would make you stop — before you have any emotional stake in the answer.

What the work actually looks like

  1. Form a hypothesis. Something specific enough to be wrong: “in the last ten minutes before a 15-minute crypto market resolves, the price systematically overreacts.”
  2. Test it against history — with realistic costs included from the first run, not added later.
  3. Try to kill it. Deliberately look for the conditions under which it fails. This is the step almost everyone skips, and it is the one that separates a strategy from a story.
  4. Deploy small, with an explicit rule written down for when you would stop.
  5. Compare live results to what the test predicted — and understand the sources of difference before concluding anything about the strategy.

That is a process, not a trick. It is closer to lab work than to gambling, and the discipline is the actual skill.

What a bot is not

It is not passive income. It is not a machine that discovers opportunities you did not have. It is not a way to trade without having a view — it is a way to apply a view consistently. And it is not a shortcut past the work of finding out whether your view is any good.

What it genuinely gives you is real: consistency, the removal of impulse decisions, the ability to test an idea on years of history in minutes, and the ability to act on rules at hours you are not awake.

Where GoGoBots fits

The five steps above are not a metaphor for how the platform works. They are the platform, so here is the honest mapping — including where it helps least.

Step 1, the hypothesis, you write as a rule rather than as code. This removes a barrier that has nothing to do with whether your idea is good, which is the only reason it matters.

Step 2, testing against history, charges costs from the first run. You supply the fee rate rather than having it read automatically, and article 1 explains why that choice is deliberate and what it costs you. Slippage is yours to set too. Both are inputs because guessing them for you would produce a number that looks authoritative and is not.

Step 3, trying to kill it, is where the real work sits and where most tooling stops. Walk-forward analysis re-tests your rule on periods it was never fitted to. Permutation testing asks whether a rule with no relationship to the data could have produced your result. Monte Carlo reshuffles the order of your trades to show what the same strategy could have done on a less kind path. None of these can prove an edge exists — nothing can. They disqualify specific ways of being wrong, and passing them means only that you have not been caught by those particular errors.

Step 4, deploying small, runs through paper, then dry run, then live, so the gap between “it works in a test” and “it works with money” gets crossed in stages rather than in one jump.

Step 5, comparing live to test, you do by re-running the backtest over the live bot’s own dates. The two will not match exactly. Understanding why is article 6, and the difference is itself a measurement rather than a nuisance.

The honest expectation to set: most of the time the report will tell you the idea does not work. That is not the tool failing. That is what the tool is for, and finding it out here costs you an afternoon instead of a balance.

The rest of this series is the methodology behind each of those steps. It gets technical, quickly.

Next: Win Rate Is Not an Edge — why the percentage of trades you win carries almost no information, and what to measure instead. That article assumes everything on this page.

Not financial advice. Past results do not predict future results. Trading prediction markets carries risk up to and including total loss of capital. Verify that trading on these venues is legal in your jurisdiction.