This article is a personal account of MEV-bot engineering for informational and educational purposes only. Statistical figures, profit estimates, and expected-value framing describe the author's research and observations and should not be read as a forecast, performance claim, or solicitation. Nothing herein constitutes investment, legal, or financial advice, nor a recommendation to engage in MEV/arbitrage activity, which carries significant technical, market, and capital risks. Conduct your own research before deploying capital.

The 98% Failure Rate — Why I'm Doing This Anyway

I'm staring at a number that should make me quit.

Ninety-eight percent. That's the estimated failure rate for arbitrage transactions on Solana. Out of every hundred attempts my bot makes, ninety-eight of them accomplish absolutely nothing. They fail. They revert. They get beaten by someone faster. They hit a price that's already moved.

Ninety-eight percent.

I found this number while digging through academic research and on-chain data, and I won't lie — my first reaction is a gut punch. I've spent months building this system, optimizing algorithms, debugging transaction construction, learning the intricacies of on-chain program execution. And now I'm learning that the fundamental reality of this game is that almost everything I attempt will fail.

So why am I still doing this?

That's the question I'm wrestling with right now, and the answer turns out to be more interesting than I expected.

The Number in Context

Let me be precise about what "failure" means here, because the word is doing a lot of heavy lifting.

A 2024 ACM research paper studying MEV bot behavior found that approximately 58% of all bot-initiated transactions on major blockchains fail. That's the general bot population — including liquidation bots, sandwich bots, and arbitrage bots of every sophistication level. For pure cyclic arbitrage — the kind I'm building — the failure rate climbs significantly higher. On-chain analysis suggests it pushes past 98%.

When I first see this statistic, my brain immediately goes to the darkest interpretation: "This doesn't work. You're wasting your time." But then I start breaking down why these transactions fail, and the picture changes completely.

Failure Is Not a Bug — It's the Architecture

Here's what's really happening inside that 98%.

Roughly 48% of failed arbitrage transactions fail because the price condition wasn't met. The bot spotted a potential opportunity, constructed a transaction, submitted it — and by the time it landed on-chain, the price had already moved. The safety check built into the transaction said "this trade would lose money" and reverted.

Read that again. The safety check reverted the transaction. This isn't a bug. This is the system working exactly as designed. It's a fail-fast mechanism. The bot tried, the opportunity wasn't there anymore, and the transaction self-destructed before any funds were at risk.

Other failure causes include write lock contention (multiple bots trying to interact with the same liquidity pool in the same slot — only one can win), stale data (the on-chain state changed between reading and writing), and plain old competition (someone else's transaction landed first with a higher priority).

None of these are engineering failures. They're market structure. The blockchain is a shared, adversarial environment where hundreds of bots are simultaneously racing for the same opportunities. Most of them must fail — there's only one winner per opportunity.

It's like showing up to an open casting call in Hollywood where three hundred actors audition for one role. Two hundred ninety-nine of them "fail." But nobody calls acting a broken profession because of it.

Bots Fail More Than Humans — And That's the Point

Here's a statistic that initially seems damning but actually reveals something profound: bots fail at a dramatically higher rate than human traders. The same ACM research shows bots failing around 58% of the time versus roughly 6% for human-initiated transactions.

A human trader looks at this and says, "See? Bots are terrible at trading."

But that's exactly backward.

A human trader submits a transaction when they're pretty sure it's going to work. They've checked the price, they've thought about it, they've decided this is a good trade. Their 94% success rate reflects extreme selectivity. They take only the shots they're confident about.

A bot operates on a completely different philosophy. It submits transactions whenever there's any chance of profit, no matter how slim. It fires hundreds or thousands of times per hour. It doesn't deliberate. It doesn't agonize. It tries, and if it fails, it tries again immediately with zero emotional cost.

This is the difference between a sniper and a machine gun. The sniper has a higher hit rate. The machine gun puts more rounds on target.

The question isn't "which one misses more?" The question is "which one accomplishes more per unit of time?"

The Asymmetry That Changes Everything

Now we get to the part that made me stop panicking and start thinking clearly.

What does a failed transaction actually cost?

On Solana, a failed transaction costs approximately $0.001 in compute fees. One-tenth of a penny. Some transactions cost even less. And with Jito bundles — the mechanism many MEV bots use to submit transactions — a failed bundle costs literally zero. Nothing. You don't pay unless you win.

Now compare that to what a successful arbitrage transaction earns. Data from on-chain analysis puts the median successful arb profit somewhere around $1.58.

Do the math. The cost of failure is $0.001. The reward for success is $1.58. That's a ratio of 1,580 to 1.

Let me say that differently: a single success pays for 1,580 failures.

At a 2% success rate, out of every 100 attempts, 2 succeed. Those 2 successes generate roughly $3.16 in revenue. The 98 failures cost $0.098. Net profit: approximately $3.06 per 100 attempts.

That's not a losing game. That's a wildly positive expected value game.

And with Jito bundles where failure cost drops to zero? The math gets even more ridiculous. Every single success is pure profit. Every failure costs nothing. The only question becomes: how many attempts can you make per unit of time?

This Is Not a Lottery

I need to be honest about a distinction that matters here, because there's a seductive trap in this reasoning.

A lottery also has a huge asymmetry between ticket cost and jackpot. A $2 Powerball ticket can win $500 million. But the expected value of a lottery ticket is deeply negative — you pay $2 and the statistical return is about $0.80. Over time, you always lose.

MEV arbitrage is structurally different. The expected value is positive. Not because of lucky jackpots, but because of consistent, small-margin wins that accumulate through sheer volume. It's not gambling — it's a probability game where the math is on your side, if you can execute at sufficient speed and volume, and if your safety mechanisms prevent you from taking losing trades.

That "if" is where all the engineering lives.

What Other Industries Already Know

The moment I reframe MEV arbitrage as a positive-expected-value probability game, I start seeing the same pattern everywhere.

Venture capital is perhaps the most famous example. About 90% of VC-backed startups fail. A venture fund expects to lose money on the vast majority of its investments. But the ones that succeed — the 10% — generate a disproportionate share of the entire industry's returns — a pattern well-documented in venture performance research. A single breakout company can return an entire fund. VCs don't succeed despite the 90% failure rate; they succeed because they've structured their game to tolerate it.

Baseball is the sport that made statistics into a religion. A lifetime batting average of .300 — meaning the batter fails to get a hit 70% of the time — is considered Hall of Fame caliber. Tony Gwynn retired with a .338 career average; Ted Williams hit .344. These are the greatest hitters in the history of the sport, and they failed roughly two out of every three times they stepped to the plate. The lesson baseball teaches is that excellence isn't about eliminating failure. It's about failing slightly less often than everyone else.

Cold calling in sales has a conversion rate between 1% and 3%, depending on the industry. That means 97 to 99 out of every 100 calls end in rejection. Yet cold calling remains a multi-billion dollar activity because the revenue from each conversion dramatically outweighs the cost of each rejection. The top salespeople don't have dramatically different conversion rates — they just make more calls and handle rejection with less friction.

Pharmaceutical R&D is another stark example. Only about 5-10% of drugs that enter clinical trials ultimately receive FDA approval, according to BIO Industry Analysis data. The failure rate across all phases is 90-95%. A single successful drug can generate billions in revenue, funding the entire research pipeline including all the failures. Pharmaceutical companies don't try to eliminate failed trials — they try to fail faster and cheaper so they can run more experiments.

Every one of these industries has internalized a truth that feels counterintuitive: high failure rates are not just tolerable but necessary features of systems that generate outsized returns on success.

Engineering the Odds

Here's where this stops being a philosophical exercise and starts being an engineering challenge.

If the baseline arbitrage success rate is roughly 2%, and the economics are already positive at that rate, then what happens if I can push it to 3%? To 5%? To 8%?

A 50% improvement in success rate — from 2% to 3% — doesn't sound dramatic. But it means 50% more revenue from the same number of attempts. In a game running thousands of attempts per day, that compounds into meaningful numbers fast.

And this is where technical skill actually matters. Not in eliminating failure — that's impossible in an adversarial environment — but in systematically shifting the probability curve.

Faster data feeds mean my view of the market is fresher when I decide to attempt a trade. Better price estimation means fewer attempts on opportunities that were never real. More efficient transaction construction means my transactions land faster and lose fewer races. Smarter opportunity selection means spending attempt budget on higher-probability targets.

Each of these improvements is marginal. None of them will turn 2% into 50%. But in a game with 1,580x cost-reward asymmetry, even small improvements in hit rate produce disproportionate returns.

This is fundamentally an engineering discipline. The market provides the asymmetry. The technology determines how effectively you exploit it.

The Failure Rate Is the Moat

There's one more dimension to this that I'm only now beginning to appreciate.

The 98% failure rate isn't just a cost of doing business — it's a barrier to entry. It's the reason most people look at this space and walk away. It's psychologically brutal to watch 98 out of 100 attempts fail, even when the math says you're winning.

Most people can't stomach it. They see the failures pile up and they quit. They don't have the patience to build the infrastructure, run the experiments, iterate on the algorithms, and trust the math over their emotions.

That psychological barrier is, in a very real sense, what makes the opportunity possible. If the success rate were 80%, everyone would do it, competition would be infinite, and margins would compress to zero. The 98% failure rate is the filter that keeps the playing field manageable.

This is the same dynamic you see in March Madness brackets. Every year, millions of people fill out brackets trying to predict the NCAA tournament perfectly. The odds of a perfect bracket are roughly 1 in 9.2 quintillion, by some statistical estimates. Nobody expects to get it right. But the engagement, the analysis, the strategic thinking — that's where the value lives. The impossibility of perfection doesn't stop anyone from playing. It's what makes the game interesting.

Of course, MEV arbitrage isn't about perfection or astronomical odds. It's about consistent execution in a game where the math works even at 2% success. The analogy is that the difficulty is the feature, not the flaw.

Starting Anyway

So here I am, staring at 98%, and I'm not quitting.

Not because I'm delusional about the odds. Not because I think I'm special enough to beat them. But because I understand, now, that the odds aren't the enemy. The odds are the game. And the game has positive expected value.

I'm building a system that will fail 98 times out of 100. And those 2 successes, over thousands and thousands of attempts, will generate returns that more than cover every single failure. The engineering challenge isn't to eliminate failure — it's to fail cheaply, fail fast, and incrementally improve the probability of success.

Every tenth of a percentage point I claw back from the failure rate is pure profit. Every millisecond I shave off latency is another race won. Every improvement in price estimation is another false opportunity filtered out before it wastes an attempt.

I know that 98% of what I try will fail. I know that this is normal. I know that this is what the game looks like when it's working correctly.

And I know that the only guaranteed way to fail at this game is to stop playing it.

So I'm starting anyway. Not in spite of the 98% failure rate — but because I finally understand what it actually means. It means this is a game where you can engineer your edge, where failure is cheap and success is valuable, and where the willingness to keep showing up is itself a competitive advantage.

That's not a reason to quit. That's a reason to build.