Nine of Ten Trading Strategies Failed: The Arithmetic That Decides Before You Fund One

In September 2026 we pointed an autonomous trading agent at a live market with a deliberately tiny budget: ten dollars. The market was a fractionalized domain token on Doma Protocol, traded through Doma's agent tooling. The goal was never to get rich on ten dollars. It was to find out which strategies survive contact with a real order book, and to write down the arithmetic that decides it.

Ten strategy families went in. Nine came out dead. The one survivor beat a do-nothing benchmark in one measurement window and lost to it in four others.

That sounds like a failure. It is the opposite. The useful output of an experiment like this is rarely a winning strategy. It is a set of numbers you can apply to any market before you risk anything, plus a clear answer to the question almost nobody asks first: how much capital does a strategy need before it is even logical to run?

For this market that answer turned out to be somewhere between five hundred and five thousand dollars, and it has almost nothing to do with how good the strategy is.

On the surface this token looked extraordinarily active. Roughly 2.5 million dollars of daily volume against a pool holding about 289,000 dollars of liquidity. That is the kind of turnover that makes a market look liquid and worth trading.

We pulled every individual swap from the chain and looked at who was actually trading. Over a fourteen day window there were 1,048,575 swaps from 3,183 wallets. Four wallets accounted for about 95 percent of all trades and 81 percent of all volume.

More telling: 94.6 percent of all trades sat inside a transaction that contained both a buy and a sell. Not a buy followed later by a sell. A buy and a sell in the same transaction, which is a round trip that ends exactly where it started.

The explanation was published on the token's own website. The token is the prize asset in a pack-opening game. Every time a player opens a pack, a contract buys tokens to back the cards and, unless the player keeps a card, sells them straight back. Four operator wallets fulfil those transactions. The game had processed over half a million pack openings, and the escrow holding kept cards contained only about 2,300 dollars of tokens. Nearly every pack was opened and cashed out.

So the 2.5 million dollars a day was not a market. It was one machine trading with itself, a million times a fortnight.

Strip out the operator wallets and the same-transaction round trips and a thin residual of genuine flow is left. Look at that residual closely and 84 percent of those trades reverse the same wallet's own previous trade within thirty minutes. Across fourteen days, only 133 trades were larger than 500 dollars.

The first lesson is the oldest one in markets, restated for on-chain data. Volume is not liquidity and turnover is not interest. Both are trivially easy to manufacture, and on-chain they are manufactured in public where anyone can check. Before treating a token as tradable, count the distinct wallets behind the volume and check how many trades are round trips inside a single transaction. Two queries, and they would have saved us a week of modelling.

Having established that the activity was artificial, we nearly stopped. One fact kept the experiment alive.

The pool charges a fee of one basis point, which is one hundredth of one percent. At our trade sizes the price impact was too small to measure. A full round trip, buy then sell, cost between two and eight basis points all in.

That is unusual enough to matter. On a typical decentralised exchange pool the fee alone is 30 basis points a side. On a centralised exchange you might pay 10 basis points a side plus the spread. A round trip costing two to eight basis points instead of sixty changes what is theoretically harvestable. A price wiggle of a fifth of one percent is noise you must ignore on most venues. Here it clears costs several times over.

This is the kind of structural detail worth hunting for, because it is the only thing that can make a small, boring market interesting. Cost is the floor under every strategy. Lower the floor and things that were impossible become merely difficult.

Price impact scaled steeply with size, and this curve sets the ceiling we come to later:

• ›A 10 dollar trade moved the price 0.01 percent
• ›A 100 dollar trade moved it 0.01 percent
• ›A 1,000 dollar trade moved it 0.15 percent
• ›A 10,000 dollar trade moved it 6.3 percent

Impact is effectively free up to a few hundred dollars, mild at a thousand, and ruinous at ten thousand.

A cheap venue is only useful if the price moves. This one barely did.

We rebuilt the price from the pool's own tick data, which removes the bounce you get from alternating buys and sells and shows the true mid price. Over the quiet period the price did not change at all in 72 percent of minutes. The median hour had a range of 0.04 percent. Three quarters of hours stayed inside 0.39 percent.

Daily ranges over six consecutive days ran 6.8 percent, 1.5 percent, 1.4 percent, 5.8 percent, 1.5 percent, and 0.02 percent. That last number is not a typo. On one full day the price moved two hundredths of one percent.

The price was pinned, and the reason was visible on-chain. Reading the pool's liquidity position and inverting the formulas showed the liquidity was not spread thinly across all prices. It was concentrated in a band of roughly plus 1.58 percent to minus 1.50 percent around the current price. Inside a band that tight it takes heavy buying or selling to move the price at all. The wallet that launched the token holds both that liquidity position and 62.8 percent of the token supply.

So the shape of this market was: very cheap to trade, and almost nothing to trade against. Movement arrived in rare bursts, then the price re-pinned.

Every idea was tested the same way. Fit it on one slice of history, then measure it on a slice it had never seen, with real costs charged. That second step is the whole game, because anything can be made to look good on the data it was built from.

• ›Resting ladder of limit orders. Beat a hold-and-do-nothing benchmark in one of five test windows. Kept as a baseline, not as an edge.
• ›Short-term momentum. Lost 4 to 5 basis points per trade out of sample. Dead.
• ›Net inflow continuation. Plus 26 basis points per trade when fitted, flat to negative when run live. Unproven.
• ›Operator buying rider. Plus 82 basis points when fitted, then the signal never fired again. Unproven.
• ›Short-term reversal. Statistically insignificant in every window. Unproven.
• ›Fading large trades. Impact measured 6 basis points, subsequent reversion measured zero. Dead.
• ›Time-of-day patterns. No effect that survived a significance test. Dead.
• ›Buying below a redemption floor. The buyout formula tracks the market price, so no floor exists below it. Dead.
• ›Cross-pool arbitrage. Minus 10 to 90 basis points per attempt after fees. Dead.
• ›Providing liquidity. Lost to simply holding at every range width. Dead, and covered in its own section below.

Three of these failed for reasons that generalise well beyond this token.

Momentum looked real and was not. Fifteen minute returns showed positive autocorrelation with a strong test statistic. Splitting the sample by day revealed the entire effect came from a single day, the one day the game operator was accumulating heavily. Outside that day it vanished. One unusual day in a short sample will happily masquerade as a tradable pattern.

The best-looking signal did not survive being run in real time. Organic net inflow above a threshold, measured in clean hourly buckets, predicted the next hour with 26 basis points of average return and a 71 percent hit rate. Encouraging. Then we re-ran it the way a live agent actually sees the world, as a rolling window updating continuously rather than in tidy hourly buckets, with realistic costs. It went flat to negative. The signal was substantially an artefact of where we drew the bucket boundaries.

Fading large trades had no edge because there was nothing to fade. The theory is that a large uninformed trade pushes the price away from fair value and you profit when it snaps back. Trades between 300 and 1,000 dollars moved the price about 6 basis points, and the subsequent reversion was approximately zero. No dislocation, no snap back, no trade.

This one deserves its own section because it is the mistake we made ourselves, in writing, and had to correct.

The reasoning looked airtight. The pool generates about 181 dollars a day in fees from all that churn, and the people providing the liquidity collect them. The churn is the ideal flow for a liquidity provider, because a buy and a sell in the same transaction pays the fee twice and leaves the price where it started. We calculated a fee yield of roughly 37 percent a year and called it the only structural edge in the market.

Then we settled the position properly. We took a 500 dollar position, replayed all 931,376 swaps in the measurement window, credited our share of every fee only while the price sat inside our range, and valued what the position actually held at the end.

• ›At plus or minus 1 percent width: 0.89 dollars of fees earned, and 1.30 percent worse than simply holding
• ›At plus or minus 5 percent width: 0.95 dollars of fees earned, and 0.42 percent worse than holding
• ›At plus or minus 25 percent width: 0.22 dollars of fees earned, and 0.09 percent worse than holding

Every width lost. Re-centring the position daily, which is what an active provider does to stay in range, made it worse, between 1.48 and 3.69 percent behind.

The arithmetic is simple once written down. Providing liquidity earns fees and loses money to rebalancing. When the price rises, the position automatically sells into the rise and is left holding less of the thing that went up. That cost scales with how concentrated the range is, and so does the fee income, so concentrating multiplies both sides of the trade equally. In this pool both sides came out at roughly 0.03 percent a day and cancelled. A one basis point fee tier simply cannot pay for the risk when the token moves one and a half percent a day.

We also checked whether the higher fee tiers rescued it, since the same token has pools at 0.3 percent and 1 percent. Over 28 hours the 0.3 percent pool traded 589 dollars and the 1 percent pool traded 156 dollars. The router only splits into them on large trades, and this market has almost none. All the churn stays in the cheap pool. Plenty of fee rate, nothing to charge it on.

Here is the part most strategy write-ups leave out entirely.

A strategy has an operating cost that does not shrink when your position does. In this system every execution is charged a minimum of one cent. A four-rung ladder re-centred once a day places four orders and cancels four, so eight executions, so eight cents a day. Call it 56 cents a week just to keep the lights on.

Now compare that fixed cost against what a one percent weekly edge would actually produce:

• ›On a 10 dollar pool, one percent a week is 10 cents against 56 cents of cost. You lose 46 cents a week before the market does anything at all.
• ›On a 100 dollar pool, one percent is one dollar against 56 cents. Costs still eat over half the edge.
• ›On a 500 dollar pool, one percent is five dollars against 56 cents. Costs are 11 percent of the edge, which is tolerable.
• ›On a 1,000 dollar pool, one percent is ten dollars against 56 cents. Costs are 6 percent. Comfortable.

Our original ten dollar budget was the right size for proving the plumbing works and the wrong size for making money, and those are genuinely different projects. Somewhere around five hundred dollars the fixed cost stops dominating. That is the floor.

The ceiling comes from the impact curve above. A ladder rung only makes sense if the impact of filling it is small compared to the distance between rungs. With rungs spaced 1.5 percent apart, a 500 dollar rung costs about 0.08 percent in impact, which is 5 percent of the rung. A 2,000 dollar rung costs about 0.3 percent, which is 20 percent of the rung. A 5,000 dollar rung costs more than the rung could ever make, and exiting a 5,000 dollar position in one go would cost about 1 percent.

So for this pool the viable window runs from roughly five hundred to five thousand dollars. Below it, fixed costs eat the edge. Above it, your own trading moves the price against you.

The critical point: that window is a property of the pool, not of the strategy. It is set by the fee, the minimum charge per execution, and the depth of the liquidity. You can calculate it before writing a line of strategy code, and you should, because it tells you whether the project is worth starting at all.

And then the honest ending. Inside that window, the measured edge was still zero to one percent a week over simply holding, and the ladder beat holding in only one of five test windows. Correct sizing makes a strategy sane. It does not make it profitable. Both questions have to be answered yes, and people usually only ask the second one.

The agent itself is deliberately boring. Orders rest as limit orders rather than market orders, both because they cost less and because a resting order cannot chase a price. A risk gate sits in front of every action that can spend money, and it is a pure function: given the proposed action and the current state, allow or deny with a reason. The strategy layer cannot place an order without passing it.

The gate refuses, among other things, to ever place a sell at or below the current market or a buy at or above it. That sounds too obvious to need stating. It exists because an earlier version of this system did exactly that once, which dumps inventory at market instantly.

The stop loss is measured against the portfolio high-water mark using a 24 hour average price rather than the last trade. It cancels everything, sells in chunks sized to fit both the per-transaction limit and whatever spending allowance remains, revokes the standing approvals so nothing can fill afterwards, then refuses to trade for a day. If it cannot sell, because the spending allowance has run out, it stops and says so loudly rather than reporting success.

That last detail came out of a review, and the review is worth describing.

We had the finished code examined by a large set of independent automated reviewers, each assigned a single dimension to attack, with every finding then handed to two further reviewers whose only job was to refute it. They confirmed 34 real defects. Five could have lost money silently.

Four separate reviewers independently found the same worst one. When a limit order filled, the system compared the order's new state against its last known state to work out what had just traded. If the exchange reported the input consumed before it reported the output received, the code advanced its bookkeeping without recording the trade, and the next comparison then booked the fill at roughly half its true price. Cost basis, profit and loss, and the reported return all went quietly wrong.

A second favourite: the ladder re-armed a replacement order one level toward the centre after each fill, except that the level at the centre does not exist, so the calculation landed on a non-existent level and gave up. Those innermost rungs are the ones that fill most often. The strategy was silently dropping most of its own mechanism, and every test passed, because the tests asserted the behaviour the author intended rather than the behaviour the market would produce.

If there is one engineering lesson here: have something adversarial read your money code, and make it verify its own findings. Reviewers with a specific dimension to attack and an incentive to disprove themselves found things no amount of staring at a diff would have.

Three numbers decide whether a strategy is worth building, and all three are available before you commit a penny.

The round-trip cost. Fee both ways, plus price impact at your intended size, plus slippage. Quote it live at several sizes rather than reading the documented fee. Ours was two to eight basis points, which was unusually good and the only reason we continued.

The fixed cost per cycle, divided by your capital. Minimum charges, gas, any per-action fee. Express it as a percentage of your pool per week. If it is a large fraction of your expected edge, you need more capital or fewer actions, and no amount of cleverness will fix it.

How often the price moves further than the first two combined. Not the advertised volatility, the real distribution. We found the median hour moved 0.04 percent and 72 percent of minutes did not move at all. If your costs are 8 basis points and the price rarely travels further than that, there is nothing to harvest and the quality of your signal is irrelevant.

If the third number is not comfortably larger than the first two, stop. You have just saved yourself the build.

Two habits earned their keep alongside those numbers. Split your sample by day before believing any signal, because if the effect lives on one or two days you have found an event rather than a pattern. And settle positions against the real price path, never against a yield estimate, which is what turned our liquidity provision result from plus 37 percent a year into a loss.

The measurement window was short. The quiet period covered six days and the token's entire history at the time was under three weeks. Every number here describes one token in one regime, and the promotional campaign driving its volume had a published end date shortly after we finished, which would very likely have changed the regime again. Conclusions about this specific market should be read as conclusions about that window. The method is what travels.

This is a write-up of an experiment, not financial advice, and nothing here is a recommendation to buy or sell anything. We were deliberately running an amount of money written off in advance, which is the only sensible way to run an experiment whose expected output is information rather than profit.