Setting Up a Copy Bot: Managing Risk and Capital

The first article in the cycle. A practical guide for someone who isn’t a mathematician and isn’t an experienced algo trader. We assume the leader has already passed screening — they’re not a scalper, not an HFT, not a market maker, not an arbitrageur, and their edge is copyable. This article isn’t about whom to copy. It’s about what to do next: how much money to stake, how to avoid getting wiped out by one of someone else’s bets, what to watch every day, when to stop, and when to add capital.

---

Someone else’s bet is not your signal

There’s one idea that reorders everything else. Grasp it, and half the mistakes of copy trading disappear on their own.

When you copy a trader, you copy their choice: which market, which side, at what price to enter. That’s where their edge lives — in the decision “there’s an advantage here.” But along with the choice, mirror copying also drags in their bet size — and size tells you nothing useful.

Why? Because the size of their bet is a function of their life, not yours. They bet big because:

• not all of their money is on this account (the deposit is only part of the capital they’ve set aside for trading);
• they have more than one account, and this is one leg of an arbitrage between them;
• they changed tactics and now bet larger;
• their deposit grew, and your system hasn’t seen it yet;
• or they simply have a different appetite for risk.

The reason doesn’t matter. What matters is the conclusion: you have no right to treat the number “how much they staked” as a signal. If you copy size one-to-one (proportionally), then the moment they go “all in,” you stake almost your entire capital on their luck. The outcome is beside the point — the point is that you’ve lost control of risk.

Hence the first principle everything else rests on:

Take the choice from the leader. Compute the size yourself.

What follows is exactly how to compute size yourself, and how to build a safety net so that one of someone else’s “all-in” bets doesn’t carry off your deposit.

---

Part 1. Proportion fixes scale, but not concentration

The first thing any platform will offer is “auto-size the bets”: they staked X% of their book, you stake X% of yours. Your money is smaller, so the same orders, just so-many times smaller. The equity curves run almost parallel, because the bets are proportional. It looks logical.

And it works — but only against one problem. Proportion solves the scale mismatch: their book is $930k, yours is $10k, the ratio is ~93, all their orders are compressed 93×. Fine.

What proportion does not solve is concentration. Watch what happens on an all-in: they stake nearly their whole bankroll, so their bet as a share of the book is ≈100%. Proportion faithfully carries that share over to you — and you stake ≈100% of your bankroll. Proportion scaled the size, but it carried the concentration over one-to-one.

The takeaway to remember: no proportional-sizing scheme on its own saves you from an all-in, because an all-in is about concentration, not scale. Concentration is cut by one thing only — a hard ceiling. And it’s placed not on the trader and not on a single market, but on a cluster of correlated bets. More on that in the safety-net part.

---

Part 2. How much to stake: two right bases (and one for emergencies)

There are two healthy logics for “how much money to stake.” Everything else is either a degenerate case of them or a trap.

First, one clarification without which both bases break. The unit you size is a position, not an individual trade. The leader builds a position not with one order but with several entries: first bought 10, then added 1,000, then another 200 — in an unpredictable order. If you compute “fixed fraction on each of their entries,” then six of their add-ons turn into six bets on one market for you, and exposure balloons several-fold — with the multiplier set not by you but by their accumulation style. So both Base A and Base B are computed from the target position in a market (which market, which side), not from the stream of their individual orders. We’ll work out exactly how to react to their add-ons inside Base A, where it’s clearest.

Base A. A fixed fraction of your bankroll (flat fixed-fraction)

Each copied position = a fixed percentage of your capital. For example, 1–2% per position — regardless of how much they staked.

In plain terms: you take only the choice from them (which market, which side, at what price), and you assign the size yourself, always the same by fraction.

When this is the right choice: when their relative size can’t be trusted. And that’s a frequent case — because their largest bets are most likely the ones where they have something you don’t: a hidden hedge on a second account, a leg of an arbitrage, an inside read on an event. Copying their “conviction” by size in those spots means amplifying a sample that works against you.

Upsides: you don’t need to know their bankroll; on a drawdown the bet shrinks by itself (bankroll down → the fraction of it is smaller in dollars); it doesn’t drop small bets below the exchange minimum as often as proportion does. Downside: you lose their “loudness” — their timid probing poke and their conviction bet weigh the same for you.

How Base A reacts to the leader’s add-ons. This is the crucial point. There’s a correct way and a simplified way; both are safe on risk, and differ in copy fidelity.

The correct way (default) — “target position.” Your fixed fraction is not a bet on each of their entries, but a ceiling on your position in that market. While the leader is building, you build behind them, spreading your fraction across their entries; once you’ve reached your fraction — you stop and add no more, however many times they keep buying. The final exposure is exactly the same as if you’d entered the full fraction at once — so on risk it’s no more dangerous than the simple option. But gradual entry has one decisive advantage: your average entry price ends up closer to their average price. And the gap in entry prices is the main leak of copy trading (you always enter a bit later and a bit worse; the closer your average price is to theirs, the more of their edge survives to reach you). Plus you don’t overcommit on a probe: if they first poked small and then changed their mind and didn’t scale up, you also stayed in a small position rather than piling into the full fraction on their one first poke.

The simplified way (if you don’t want to run step-by-step accumulation) — “first entry only.” You stake your fixed fraction when the leader first enters a market, and you ignore all their add-ons in the same market. Simpler (no need to track accumulation), equally safe on exposure (one fraction per market). The downside is worse entry-price matching (you take the whole size at the price of their first poke) and the risk of piling into the full fraction on a small probe they didn’t follow up. Serviceable as a default mode for a beginner; all else equal, “target position” is more precise.

In both ways, the last word belongs to the cluster cap (next part). It’s the wall your accumulation stops against. Without a cluster cap, a “fixed fraction” with add-ons has no upper bound on risk. Remember the pairing: Base A sets the size of one position, the cluster cap limits the sum of correlated positions — they work only together.

Base B. Normalizing to their conviction, but with a clamp (winsorization)

They staked X% of their book → you stake X% of your bankroll. This preserves their relative conviction: where they bet larger, you bet larger too.

But in raw form this is exactly the variant that carries an all-in over to your ~100%. So we do not take it raw. We take it with a clamp: we limit how “loud” their largest bet can be relative to their usual one. In practice — we clamp their multiplier relative to their own median bet to about 2–3×. That is, if they usually bet “one unit,” their largest bet in your translation counts as at most “two or three units,” not “thirty.”

Simply put: you hear their conviction, but you don’t let them shout at you.

When this is the right choice: when their relative size is a real signal (they genuinely bet larger where the advantage is stronger), and that signal carries over to you.

Base B has no separate add-on headache: you copy their final position share, not the stream of entries, so their six add-ons simply converge to a single share for you — the ballooning seen in Base A doesn’t arise here by construction. (This, incidentally, is another argument for framing any base through the target position rather than through individual trades.)

How to choose between A and B — without formulas

(Base C — a fixed dollar amount, described right after this — isn’t part of the choice: it’s a forced measure, only for a small bankroll. The choice is between A and B.)

The choice is decided by one test, not by market type and not by the leader’s “experience.” The question is simple: does the size of their bet carry useful information?

Here’s how to check. Split their history into groups by bet size — small, medium, large. For each group, compute whether it was profitable on average from the copier’s point of view (that is, accounting for the fact that you enter later and at a slightly worse price, and pay your own fees — more on this below). Then two outcomes:

• Profit rises with size (the leader’s large bets are on average better than small ones) → their size = conviction = useful → take Base B with a clamp.
• Profit is unrelated to size, or large bets are actually weaker → their size = noise or hidden structure → take Base A, the fixed fraction.

Until this test is done, the choice of base is unjustified — it’s just guessing.

Base C. A fixed dollar amount — only for a small bankroll

There’s a third variant: staking a fixed amount (for example, “always $200 per position”) rather than a fraction. It has a narrow but real use case — when your bankroll is so small that a fixed fraction produces an order below the exchange minimum. Then a fixed amount is a forced measure to get the order to go through at all: you raise the size to the executable minimum, sacrificing “correctness” so that copying physically works.

In all other cases a fixed amount is worse than a fixed fraction, and here’s why: on a drawdown, a fixed amount turns into a growing share of the bankroll. Lost half — your $200 is now 4% of the bankroll instead of 2%, and you’re adding risk precisely when you’re losing. That’s the opposite of sound management — a fixed fraction, by contrast, shrinks the bet by itself on a drawdown.

Bottom line on bases: by default — a fixed fraction (A) or normalization with a clamp (B), with the choice between them decided by the test above. A fixed amount © — only when the bankroll hits the exchange floor.

And what definitely not to do

Copying their absolute one-to-one — with your smaller bankroll this is physically impossible; any attempt to “scale the absolute” is just the normalization from Base B.

Placing a fixed fraction on each of their fills (rather than per position) — brings back exactly the add-on problem for which we introduced “the unit = position.” Don’t confuse an entry with a position.

---

Part 3. The safety net: four circuit breakers

The base answers only “how much money per bet.” It does not answer what happens when several bets turn out to be one big bet, or when the leader breaks. So on top of any base — A or B — four circuit breakers are placed, without exception. This isn’t an option for the “cautious.” It’s part of the construction.

Breaker 1. Fractional Kelly — “don’t go all in even when you’re sure”

This needs explaining from scratch, because the word makes many people’s eyes glaze over, though the idea is actually simple.

What it is and where it came from. The “Kelly criterion” is a rule that answers one question: if I have an advantage, what fraction of the bankroll should I stake so that capital grows fast and I still don’t go broke? It was derived in the 1950s by John Kelly at Bell Labs — precisely for repeated bets with an advantage. The idea in plain terms: bet too little and money grows slowly; bet too much and one losing streak zeroes you out; between these extremes there’s a sweet spot, and Kelly computes it.

How the number is computed. For a bet “buy an outcome at price c that wins with probability p,” the rule gives the fraction of the bankroll:

f = (p − c) / (1 − c)

Decoded in human terms: the numerator (p − c) is your advantage, i.e. how much your probability estimate exceeds the price you pay. The denominator (1 − c) is what you risk per dollar. If there’s no advantage (your estimate equals the price, p = c), the numerator is zero → f = 0 → don’t bet at all.

A small example, to defuse the fear that “Kelly will make me bet big.” Say an outcome costs c = 0.50, and you estimate its chance at p = 0.55 (a 5-percentage-point advantage — already a decent edge for such markets). Then f = (0.55 − 0.50) / (1 − 0.50) = 0.05 / 0.50 = 0.10, i.e. 10% of the bankroll. And that’s full Kelly on a good edge — already not “half the deposit.” On a more realistic advantage of 2–3 points, f comes out in the low single-digit percents. So even without any clamp, Kelly on its own doesn’t push you into an all-in.

Why in copy trading it can’t be computed exactly — and what we do instead. The formula needs your estimate of probability p. And in copy trading you don’t know p: you only see that the leader entered, and at what price. The entry price is the market’s estimate, not your estimate of the advantage. There’s nothing to feed the formula. So Kelly works for us not as a calculator for each bet, but as a justification that a ceiling should exist at all, and roughly at what level: you know the order of magnitude of the copyable slice’s advantage (low single-digit percents — from analyzing the leader), and Kelly says that a reasonable bet on such an edge is also low single-digit percents of the bankroll, not tens. The real size limiter is the cluster cap (Breaker 2); Kelly sets the scale, the cap is the hard wall.

Why a quarter or a half, not full (“fractional Kelly”). Full Kelly assumes you know p exactly. You don’t — you estimate it with error, and indirectly, through someone else’s actions. The rule is simple: the less sure you are of your estimate of the advantage, the smaller a fraction of Kelly you take. “Fractional Kelly ½” literally means “I act as if my advantage were half of what it seems” — insurance against overrating yourself. A practical bonus: half-Kelly loses very little growth speed but noticeably reduces drawdowns. In copy trading, where p is estimated crudely and through an intermediary, a reasonable band of caution is ¼–½ (that is, we take between 25% and 50% of what full Kelly would compute). Full Kelly here is almost always “too bold.”

Bottom line: Kelly is the dial “if in doubt, bet less,” which sets the right order of magnitude for the bet. On its own it neither sizes each trade nor protects against concentration — that’s the next breaker’s job.

Breaker 2. The cluster cap — the main shield against an all-in

This is the most important breaker, and also the most misunderstood, so let’s go through it in detail.

What a cluster is. A cluster is a group of bets that actually win and lose together, because they depend on one and the same event. A simple example from politics: the market “candidate X wins,” the market “X takes state A,” the market “X takes state B,” the market “party X gets the majority” — formally four different markets, but in substance this is one bet “X wins,” broken into pieces. If X fails, all four lose at once. The same in any category: several bets on one team across different props of one match — one cluster; several temperature “buckets” for one city and day — one cluster; several markets on one macro release — one cluster; several “BTC up” bets in one time window — one cluster. So risk must be counted as one bet, not as several independent ones.

What the cap does. The cluster cap limits how much money can stand on one such correlated group at a time. A starting reference is 3–6% of your bankroll per cluster of directional risk. Hit the ceiling and no new correlated bets open into that cluster, however many the leader spawns.

Why the ceiling is placed on the cluster, not on a single trade. This is the key. Imagine you set a limit “no more than 2% per trade.” The leader goes all in on one idea — say, “X wins.” They can’t push through your limit with a single bet — but they open five “different” markets on the same idea (overall outcome, individual states, majority, and so on). Each is 2%, each within your limit, all “by the rules.” And in aggregate — 10% of the bankroll on one event, i.e. your limit is bypassed by fragmentation. A per-trade limit is bypassed by spreading; a per-cluster limit is not. That’s why the cluster cap is what actually catches the “they went all in” scenario — whether via one large bet or five small ones on one idea.

How a cluster is identified — by fact, not by label. You can’t group by market name (“these are different markets after all”). You group by whether the positions move together: is it one direction, one time window, one underlying asset or event. A practical sign: if two markets almost always win and lose in sync — it’s one cluster, however differently they’re named.

Connection to add-ons (Part 2). Remember the “target position” way? The cluster cap is exactly the wall your accumulation stops against. The leader keeps buying and buying — you follow until you hit the cluster ceiling, and there you stop. Without this ceiling, a “fixed fraction” with add-ons has no upper bound. So Base A and Breaker 2 aren’t two separate mechanisms but one pairing: the base sets the position size, the cap limits the sum of correlated positions.

Important: the cap is always computed as a percentage of your bankroll, not through an estimate of theirs. Then an error in estimating the leader (for example, if their deposit grew and you didn’t notice) doesn’t breach your defense — the ceiling stands on your side.

Breaker 3. The anomaly guard — “if they suddenly bet 10× their usual, don’t follow blindly”

Look not at the absolute size, but at how much the bet deviates from their own norm. If they suddenly bet 10× their usual (median) bet — that’s a flag: a change of strategy, a leg of an arbitrage, or an error. The reaction to the flag is not “faithfully mirror it,” but one of: clamp harder than usual, pause until confirmation, or skip. You don’t need to work out why they did it. You need to limit it.

Breaker 4. The execution floor and the kill-switch

Floor: the exchange won’t accept an order smaller than a certain minimum. If your bet falls below it after all the clamps — the trade simply doesn’t go through. This is a separate problem (you lose part of their decisions from the bottom); we’ll return to it when choosing capital.

Kill-switch: if the entire copy allocation has drawn down by a preset amount — a reference is −15% to −20% — copying stops automatically. This is the last line for when the leader has genuinely broken and the anomaly guard reacted too late.

The full formula for one bet’s size looks like this (in words, left to right):

base (A or B), set at the scale fractional Kelly justifies → clamp with the cluster cap → check it’s not below the exchange minimum → and above it all hangs the kill-switch on total drawdown.

Remove the cap and the kill-switch, leave bare fractional Kelly — and you’re again exposed to exactly the all-in you started from. Kelly manages the size of one bet. It knows nothing about five legs being one direction, and it doesn’t catch a losing streak. Capital management is “in play” when all four breakers are in place, not when Kelly alone is connected.

---

Part 4. What data you need from the leader before turning the knobs

You can’t configure a copy bot “by eye.” The parameters above (which base, which cap, where the stop) are conclusions drawn from the leader’s data, not numbers off the top of your head. Here’s what to gather and what to look at. None of it requires mathematics — just reading tables.

The distribution of their bet sizes. What their typical bet is (the median), how spread out the sizes are, how often outliers occur. From this comes the anomaly-guard threshold (how many times over is “too large”) and the clamp for Base B.

The distribution of entry prices. At what prices they enter. This is critical, and here’s why. If they mostly enter at “near-certain” prices (0.90 and up), they’ll have a high win rate — but the premium there is thin, and when you enter a bit later and a bit worse, it vanishes. A high win rate on its own is not good news: it’s often the profile of a seller of “near-sure” outcomes, whose profit lives on a narrow strip of prices. The most copyable zone is mid prices (roughly 0.3–0.7): there’s both an advantage there and headroom for your slightly-worse entry. Both price edges (near 0.95 and near 0.05) are fragile for the copier.

The result of the “profit by bet size” test. The same test from Part 2. It decides the choice of base. Without it — don’t configure.

The speed of edge realization. How fast their bet “plays out.” If they enter and the price converges toward their being right over days — you have time to enter at a close price, the edge is copyable. If everything resolves in seconds to minutes (frequent fast fills) — you’re always late, and the lag eats the advantage. A practical proxy is how many trades they do per unit of time and over how many “steps” a position is assembled. “The position is assembled in a couple of entries” is a good sign; “dozens of entries per minute” is a bad one.

An estimate of their bankroll (from turnover). Needed only for Base B, to know the denominator “X% of their book.” Remember: this estimate lags. That’s exactly why all your ceilings are computed as a percentage of your bankroll, not through theirs — then an error in estimating their bankroll isn’t fatal.

The structure of their portfolio — what’s tied to what. Not by market label, but by fact: which of their positions move together. From this the clusters for the cap are identified. Several directional positions on one outcome or one event (in any category — an election, a match, a weather day, a macro release) are one cluster, not several independent bets.

Whether their PnL is “clean.” Make sure their profit and win rate come from trading, not from accruals: interest on collateral, referrals, liquidity rewards. These inflows are always “in the plus” and inflate both profit and win rate, creating an illusion of skill. They must be subtracted before you assess them as a trader.

---

Part 5. Setting the parameters: a concrete example on $10,000

Let’s put it all together. The numbers are illustrative — real ones are calibrated on the leader’s data.

Suppose the leader passed screening, and the “profit by size” test showed their large bets are on average better than small ones (size is informative) → take Base B with a clamp. Their median bet is, say, 1.5% of the book; occasionally they bet 60% and more (that’s their all-in).

• Conviction clamp: clamp their multiplier over the median to 2–3×. Their 60% bet in your translation counts not as 60% but as ~3–4.5% (2–3× of 1.5%).
• Fractional Kelly: not a separate multiplier here — it’s the reason the two numbers above are set in low single-digit percents rather than tens. As Breaker 1 says, with p unknown Kelly sets the scale, it doesn’t size each bet.
• Cluster cap: 5% of the bankroll = $500 per correlated group.
• Kill-switch: −18% of $10,000 = copying stops at a drawdown down to ~$8,200.
• Floor: an order doesn’t go through below the exchange minimum (verify at the source).

How it works in three situations:

The point is visible at once: pure proportion stakes thousands of dollars on one of someone else’s all-ins — the thing you feared. Our scheme preserves the gradient in the body (the $30 probe is less than the $300 large bet — you still hear their conviction) and clips the tail to a safe level.

You deliberately break the correlation with their extremes: on ordinary bets you go with them, on their all-ins you diverge. That cuts off both tails of the distribution at once. The lower tail you close off is the gain — their blow-up no longer kills you. The upper tail you give up is the price — you catch their home run only partly. The trade is right because past PnL is a weak predictor of the future and the size signal is contaminated by hidden structure: on average that upper tail is worth little, and the lower one is your survival.

---

Part 6. Running it: what to watch every day

The copy bot is live. Now it can’t be left unattended. Here’s the dashboard — what to look at and what each number means.

Realized copy-slice return versus the leader. Not “their” return, but yours — accounting for the fact that you entered later and paid your own fees. It should stay in a reasonable relation to the bulk of their result. A systematic gap downward is an alarm.

How many of their decisions you actually catch (coverage). If the exchange floor cuts a lot of small bets (in practice it’s happened that under strong compression 70%+ of their decisions were lost), you’re copying not them anymore but their large legs — and that’s a different sample with different behavior. Watch the share of missed decisions.

Entry quality (drift). How much worse your entry price is than theirs. A small gap is normal (you’re always a bit later). A growing gap means edge is leaking into lag; the leader may have sped up.

Fees as a share of the result. Small frequent bets can be eaten by fees. If the fee load is growing relative to profit — a signal that the sizing scheme is dropping you into orders that are too small/too frequent.

Drawdown versus the kill-switch. Always see how close you are to −15% to −20%. This isn’t for panic — it’s for discipline.

Cluster loading. How often you hit the cluster cap. If it’s constant — the leader concentrates heavily and you’re systematically under-copying. This isn’t a malfunction, but it’s information for the capital decision (Part 8).

The “cleanness” of their result over time. Periodically recheck that their profit still comes from trading and hasn’t started arriving from accruals/rewards. A change in the source of income = a change of archetype = grounds to reconsider copyability.

---

Part 7. When to stop

There are hard stops (they fire automatically) and soft warnings (they require your decision).

Hard stop — the kill-switch fired. The drawdown reached the threshold. Copying is stopped. Don’t “buy the dip” and don’t disable the kill-switch to “give them a chance to recover.” Once you’ve stopped — you figure out what happened, and only then decide whether to restart.

Soft warnings — grounds to stop manually:

• The leader changed regime. Their frequency, market categories, size distribution, or entry-price distribution changed noticeably. The person you selected is a set of habits. The habits changed — the selection is no longer valid, and re-evaluation from scratch is needed.
• The copy-slice edge thinned out. Your realized return (with your fees and your lag) has slid toward zero or negative, even if the leader is still in the plus. That means the copyable part of their advantage is exhausted.
• Adverse selection set in. You increasingly catch their losing bets and increasingly miss the winning ones (for example, their best trades are the fastest, and you don’t make it into them). This is a silent killer: the leader is in the plus while the copier is in the minus.
• Their “plus” turned out to be accruals. A recheck showed the profit came from interest/referrals/rewards, not from trading. There’s nothing there to copy.

The general rule: you stop on your data, not on their mood. Their chart may be green while your copy-slice is red. Yours decides.

---

Part 8. When to add capital

Here the most frequent and costly mistake is made: “they’re up — I’ll add more.” Past PnL is a weak predictor of the future (in our runs the correlation of past result with future is on the edge of noise). “They’re up” is not a reason to add. The reason to add is that your copy works as intended, over a sufficient sample, and that the extra capital genuinely adds something.

You can add capital when all of the following hold at once:

• A sufficient out-of-sample. Your copy bot has run not for “a couple of days” but for enough trades, and over them your realized copy-slice is confidently positive. Calendar time without a trade count means nothing.
• Positive EV specifically for you. Not their profit, but yours — after your fees and lag.
• Headroom against the caps. If you constantly hit the cluster cap, extra capital partly sits idle — first it makes sense to understand whether a higher cap is warranted, rather than just pouring money in.
• The floor stops being a problem. If you used to lose many decisions to the exchange minimum, then with more capital those decisions start going through — and here adding really increases coverage, not just bet size. That’s a good reason.

And above all — add in a ladder, not one leap. Raise capital one step → let it run over a new sample → check the metrics hold → the next step. A sharp jump in capital is essentially a new experiment without confirmation.

Symmetrically: if your realized EV slides, the caps constantly cut, and coverage falls — reduce capital by the same ladder and logic.

---

Part 9. Checklist of the data you need to get from the leader

In summary — what to pull and keep on hand for a full assessment of how effectively the copy bot works. This answers the question “what data to get from the leader.”

About the leader themselves (for setup):

• the full trade history (entries/exits, prices, sizes) — not a snapshot of positions, but the activity itself;
• the distribution of bet sizes (median, spread, outliers);
• the distribution of entry prices;
• a breakdown by market category and horizon;
• trade frequency and the number of entries to assemble one position (edge-realization speed);
• the event types in their activity — to separate trading from accruals/rewards and from structural operations (conversions, splits/merges) that can’t be copied;
• a reconstructed “clean” PnL from trading (with accruals subtracted), not the raw number off the leaderboard;
• the structure of position overlaps (what moves together → clusters).

About your copy (for maintenance):

• your realized copy-slice PnL (with your fees and lag);
• the share of their decisions caught / missed (coverage);
• your entry drift versus their price;
• fee load as a share of the result;
• current drawdown versus the kill-switch;
• how often you hit the cluster caps.

The reading rule: look at their activity, not at a snapshot of positions. A position snapshot will readily show an “ideal candidate”; the activity reveals how they actually make money. And everything is measured in trades, not in calendar days — 30 “quiet” days say nothing, 300 trades do.

---

What to verify at the source before launch

These parameters change and are not asserted from memory — verify them at docs.polymarket.com and in the current rules as of the launch date:

• the minimum order size (the execution floor and how many decisions you lose from the bottom depend on it);
• the taker fee and the exact fee formula (they enter into your copy-slice EV calculation);
• the rules on self-matching / wash trading in the Code of Conduct — if you run several copy bots on one leader, they can cross with each other, which is treated as wash trading regardless of intent; architectural self-match prevention is required, not concealment.

The starting references for the risk parameters (a cap of 3–6% per cluster, fractional Kelly ¼–½, a kill-switch of −15% to −20%, a conviction clamp of 2–3× the median) are a starting point, not laws. Final values are calibrated on the specific leader’s data with the same “profit by bet size” test.

---

One-page summary

1. You copy the choice, not the size. Their number “how much they staked” is not your signal.
2. You compute the size yourself, and you size a position, not an individual entry (otherwise the leader’s add-ons balloon exposure several-fold). Two bases: a fixed fraction of your bankroll (when their size can’t be trusted) or normalization to their conviction with a clamp (when their size is informative); a fixed dollar amount — only if the bankroll hits the exchange minimum. The choice between A and B is decided by the “profit by bet size” test, not by market type or “experience.”
3. On top of the base — four breakers, always: fractional Kelly, the cluster cap (the main shield against an all-in), the anomaly guard, the floor + the kill-switch. Kelly alone ≠ capital management.
4. You configure from the leader’s data, not by eye: sizes, entry prices, realization speed, PnL cleanness, cluster structure.
5. You maintain by the dashboard: your copy-slice, coverage, drift, fees, drawdown, caps.
6. You stop on your data, not on their mood: the kill-switch, a change in their regime, the thinning of your edge, adverse selection.
7. You add capital only on a confirmed sample and only in a ladder — “they’re up” is not a reason.

All of this is written to be checked, not believed. Verify every number on your own data and at the source.

---

Disclaimer. This is not financial advice. Past results do not predict future ones. Copy trading carries risk up to the total loss of the funds committed. Check whether such activity is permitted in your jurisdiction. Your funds remain in your own wallet at all times.