Scalping Like a Remora: Don't Beat the Big Player, Ride the Flow They Leave Behind

1. Starting point: what did Meruem "learn" against Netero?

If you had to pick one machine-learning term for how Meruem broke through Netero's Guanyin Zero, the closest is online opponent modeling. It means watching your opponent during the fight and slowly updating your guess about what they will do next in each situation.[1]

Calling it plain reinforcement learning is too rough. Meruem does several things at once:

1. He watches Netero's last move, distance, attack direction and rhythm.
2. He works out Netero's habits, the biases Netero doesn't know he has in each situation.
3. He predicts the next move.
4. He changes how he attacks, so that Netero gives a certain response.
5. He searches a few moves ahead for a way through.

So the most accurate label is opponent modeling + online learning + sequence prediction + planning. Think of it as Meruem taking the skill he built in Gungi, reading the other player's choices and comparing branches ahead, and using it in a fight.

2. Even a one-on-one fight has too many patterns. Why can it still be read?

Say each side has only five choices: attack, defend, dodge, move, feint. One turn for both already gives 25 combinations. Ten turns give 25^10, about 95 trillion. In a real fight, direction, speed, distance, strength and timing change smoothly, so the real number is close to endless.

Meruem did not memorize every sequence. What matters is that people do not use all their options equally often.

Say Netero has 100 ways to defend. At one distance he picks A 50% of the time, B 30% and C 15%, and almost never the rest. Then you only need to study A, B and C closely. If Meruem also attacks in a way that makes A more likely, the choices shrink even more.

So the point is not to try every pattern. It is to:

Narrow the situation, find the opponent's bias, and use your own moves to limit how they respond.

3. In the market, you can't see who the "strong player" is

In the duel there is one opponent, Netero, so you can learn his habits. A market has big investors, brokers, market makers, high-speed traders, discretionary traders and small day traders all at once. From the normal order book and trade list, you can't tell whose order is whose.

That makes "copy the strong trader" a risky idea for scalping. You never see the person. You only see the marks their trades leave behind:

• How many ticks the price moved after a buy order.
• Whether the price failed to fall even though people were selling.
• Whether sell orders that were eaten up were quickly refilled.
• Whether buy orders moved up to a higher price.
• Whether trades kept happening after a new high.
• Whether the spread stayed narrow or suddenly widened.

So the market version of the question is not "who is buying?" It is:

In this situation, which side, buyers or sellers, is really moving the price?

4. Research that suggests a remora-style approach can work

A remora is a fish that swims along with a shark and eats what is left. The idea is "don't race the strong player who moves first. Take what is left of the flow after them." A few studies back this idea.

First, big orders get split up. If you place a huge order at once, you push the price against yourself, so big investors often cut a large parent order into small child orders. A study of more than nine years of account-level data from the Tokyo Stock Exchange found support, in numbers, for the claim that order splitting explains why a buy tends to be followed by another buy, and a sell by another sell, for a long time.[2]

Second, other traders read split orders that are easy to predict. The clearer the pattern of a big order's child orders, the more other traders predict it and trade ahead of it, and the higher the cost and the later price impact for the big investor.[6][7] This is close to "back-running" or "order anticipation": guessing the next order from past order flow and following it.

Third, the very first race is too fast for people. A study of UK stock exchange messages found that a typical race decided only by speed lasts about 5 to 10 microseconds, and more than 20% of trading volume was linked to some kind of speed race.[9] A person looking at a screen and clicking cannot realistically compete for that first move.

So what a person can aim for is not a price gap that disappears in microseconds. It is split orders, follow-up orders and thin liquidity that last seconds to minutes.

But this does not mean "buying the second wave after a surge always wins." When a big order is done, part of its price impact can reverse, and the first buyer may sell to the later ones and leave. The core of the remora plan is not the fact that the price rose. It is checking that the flow has not ended yet.

5. Four traces you can see in public data

5-1. Order flow imbalance (OFI)

OFI counts not only market orders but also limit orders added or cancelled at the best prices, and measures the difference between buying pressure and selling pressure. A study of 50 US stocks found that short-term price changes were tied more steadily to OFI than to plain trading volume, and that the same OFI moved the price more when the order book was thin.[3]

Think of a tug of war. It's not just how many people joined. It counts which way the rope really moved, and whether people on the other side joined or dropped out.

5-2. Order book imbalance (queue imbalance)

Call the quantity on the best buy side B and on the best sell side A. A simple measure of the imbalance is:

QI = (B - A) / (B + A)

The closer QI is to 1, the thicker the buy side looks. The closer to -1, the thicker the sell side. A study of 10 NASDAQ stocks found a statistical link between QI and the direction of the next mid-price change, with the biggest gains in prediction for stocks with large tick sizes.[4]

But a thick order may not stay. If it is cancelled, it disappears. So don't only look at a snapshot of the thickness. Watch how adds, cancels and trades change over time.

5-3. How the price reacts to orders

The most useful thing is also the easiest to see. More than the size of the orders, look at how the price moved when it took them.

• A lot of buy market orders came in, but the price didn't rise: strong selling above may be absorbing them.
• Sell market orders came in, but the price didn't fall: buyers below may be absorbing them.

If you plan to follow the buyers, a stronger sign than "there are many buyers" is that the price holds even after sellers hit back.

5-4. How fast the order book recovers

After a big order eats into the book, how quickly the book fills back up is called resiliency. A study of one stock on the London Stock Exchange found that the book did not always recover, but when it did, the half-life was about 20 seconds. It also reported signs of follow-up market orders.[5]

This is not a universal rule like "trade in 20 seconds." Speed differs by market, stock and period. What matters is to measure, in the short time after the first move, whether the other side of the book comes back or not.

6. Entry idea: don't buy the first move, buy the second push after it held

The simplest remora idea on the buy side has five steps:

1. First move: price, trade frequency and positive OFI jump above normal.
2. Wait: don't chase. Wait for sellers to push back.
3. Check it holds: sell market orders come in, but the pullback is shallow, buy orders get refilled, and the starting point of the first move doesn't break.
4. Second push: positive OFI grows again, sell orders get eaten, and the price makes a new recent high.
5. Join: take only the short run that continues after the second push is confirmed.

Example in numbers: the price jumps fast from 1,000 yen to 1,010 yen, pulls back to 1,007 yen, and when sellers hit it, it doesn't break 1,007 yen, then goes above 1,010 yen again. That point is a candidate. You give the first 10 yen to the strong player and take only what is left after confirming.

However, this is not a sure-win rule finished in a paper. What research shows directly is that order direction persists, that OFI and book imbalance can predict a little ahead, and that people do trade ahead of predictable orders. The five steps above are a hypothesis to test, built for individuals from those findings.

7. Take profit by how long your reason lasts, not by how many yen you want

The reason to buy with this method is this belief:

"Short-lived buying momentum is still there, and followers will buy soon."

So it is natural to decide the exit by whether that belief is still alive, not only by how much profit you have.

Think of profit-taking and exits as four kinds, used together.

7-1. Exit when the reason breaks

These changes are signs that "buying will continue" is no longer true:

• The price breaks back down right after breaking out.
• The price breaks the low of the pullback.
• OFI turns toward selling.
• Buy market orders keep coming, but the price doesn't rise.
• Sell orders keep being refilled quickly.
• Buy orders run away and the spread widens.

In these cases you exit even if you haven't reached your profit goal. If you were running behind a strong runner and the runner stops, there is no reason to run alone to the distance you picked at the start.

7-2. Time stop

This method aims for a flow that continues right away. If no new high or second push comes within the time you expected, nothing happening is itself information.

But there is no right answer like "20 seconds for every stock." Each signal fades at its own speed, and a model of dynamic trading shows that when there are costs, you should adjust your position to how long the signal lasts.[12] From your own data, measure how many seconds or events it takes for winning trades to start rising, and use that shelf life as your time stop.

7-3. Fixed target

A fixed target is not bad, but "3 ticks" or "win 2 for every 1 lost" with no basis is weak. Collect the maximum favorable excursion (MFE, how far the trade went your way) after entry, subtract fees and slippage, and compare which target gives the best expected value.

Simply put, expected value is:

expected value = win rate × average win − loss rate × average loss − trading cost

A farther target gives more per win, but you reach it less often. Look at the size of the gain and the chance of reaching it together.

7-4. Trailing stop and taking profit in parts

To keep the rare trades that run far, you can use a trailing stop, which raises your exit line as the price rises. Math research shows that, if you assume a certain kind of price process, a limit sell order combined with a trailing stop can be the best choice.[11] But that doesn't mean "always 2 ticks below" is best in every market.

Taking profit in parts tends to make results less bumpy, but you let go of the profit on your best trades early, and you add more trades and more cost. There is no general evidence that "sell half" is always better. You need to compare a fixed full exit, a time stop, a trailing stop and partial exits on the same data.

8. A tighter stop-loss is not automatically better

A theory study on stop-loss found that if prices follow a random walk, a mechanical stop lowers expected return, but if returns tend to continue in the same direction, that is, there is momentum, a stop can be worth having because it cuts losses.[10]

The remora method aims at short-term momentum continuing, so a stop at "the reason broke" may suit it. But a stop tighter than ordinary small wiggles gets hit again and again, and all you pay is the spread and fees.

So it makes sense to split the roles:

• Initial stop: the price or order state where the setup no longer holds.
• Time stop: when the period in which the flow should continue has passed.
• Stop after profit: follow the latest pullback low or a reversal in order flow.

9. People are biased when they take profit

A well-known study of about 10,000 accounts found that investors tend to sell winning stocks too early and hold losing stocks too long. This is called the disposition effect.[13]

Scalpers fall for the reverse too. You don't want to lose a 2-tick gain, so you take it right away, but you wait for a 2-tick loss to come back. The result is small wins and big losses.

To avoid it, don't exit because it feels safe. Exit for a reason you set in advance:

• The reason broke.
• The time ran out.
• The statistical target was reached.
• The trailing condition was hit.

10. What research has shown, and what it has not

Fairly well supported

• Opponent modeling, which estimates how others act and changes your own policy, is a working research area for problems with many players.[1]
• Splitting big orders is one cause of orders continuing in the same direction.[2]
• Short-term price changes are tied more strongly to OFI than to volume alone.[3]
• Book imbalance can sometimes statistically predict the next price direction.[4]
• Other traders reading predictable big executions and trading after them, called order anticipation, has been studied in both theory and data.[6][7]
• The fastest price races are too fast for humans to join.[9]
• The value of stops and trailing stops depends on the price process, how long a signal lasts, and trading costs.[10][11][12]

Not generally proven yet

• That buying a shallow pullback and a re-breakout after a surge lets an individual win steadily after costs.
• That you can identify the orders of "skilled scalpers" by looking at the book.
• That 20 seconds, 3 ticks, or selling half is the best value for all stocks.
• That high prediction accuracy turns directly into high net profit.
• That a rule that worked in one past period will keep working after the players and the rules change.

11. A minimum design for testing it

Before handing it all to an AI, it is better to first record one kind of situation 100 to 300 times.

Record of the situation

• Stock, time, spread, tick size.
• Size and duration of the first move, and trade frequency.
• OFI, or a simple measure of buying and selling pressure.
• Size of the pullback after the first move.
• How far the price fell when sellers hit.
• Adds, cancels and refill speed of buy and sell orders.
• Whether a second push and a new high happened.

Record after entry

• Profit and loss after 1, 5, 10, 20 and 30 seconds.
• Maximum favorable excursion (MFE).
• Maximum adverse excursion (MAE, how far the trade went against you).
• Time until the highest price.
• Time until OFI reversed.
• Actual fill price, slippage and fees.

Exits to compare

1. Fixed target and fixed stop.
2. Reason breaks + time stop.
3. Fixed target, then a trailing stop on the rest.
4. Exit everything when order flow reverses.

Split the learning period and the testing period by date. Don't mix in future data. Build the rules on older data and check them on newer data. Because the market changes, you also need to repeat the test over time (walk-forward testing).

Last, always compare after costs. Even if your forecast is right, if the gain is 1 tick and the spread plus slippage is 1 tick, it is not a working strategy.

Conclusion

The main idea from Netero versus Meruem is not to memorize every pattern.

Learn the bias from behavior you can see, narrow the situation, check how the other side reacts, and then enter the short path that is still open.

In the market, this means you don't look for one named "super trader." You read the marks that strong orders leave in public data. Leave the first move to the fast players, and follow only when the pressure of orders speeds up again after it survived a counter-push.

For taking profit, don't ask only how many yen you made. Ask:

Is the reason to follow still alive? Did it continue within the time you expected? Is there still room to move after costs?

In the shortest form:

Don't try to beat the leader. Check that the flow the leader made is still going, ride it briefly, and when it stops, get off no matter what distance you aimed for.

This is the basic shape of "remora-style scalping" that you can reasonably draw from the research.

---