1. Starting point: What kind of “learning” did Meruem use against Netero?
If we must choose one machine-learning term, Meruem’s method is closest to online opponent modeling. During the fight, he observes Netero and continually updates an estimate of which move Netero is likely to choose in each situation.[1]
Calling it simple reinforcement learning is too vague. Meruem is doing several things together:
- Observing distance, direction, timing, rhythm, and recent actions.
- Estimating unconscious biases in Netero’s choices.
- Predicting the next action.
- Changing his own attack to invite a particular response.
- Searching several steps ahead for a path through the defense.
A better description is opponent modeling + online learning + sequence prediction + planning. His experience in Gungi trained him to read an opponent’s preferences and compare future branches, and he transferred that ability to combat.
2. Even a one-on-one fight has too many patterns. Why can it still be read?
Suppose each fighter has only five choices: attack, defend, evade, move, and feint. One simultaneous turn already has 25 combinations. Ten turns produce 25^10, roughly 95 trillion possible sequences. Real combat also contains continuous differences in angle, speed, distance, force, and timing, so the true space is practically unlimited.
Meruem does not memorize every sequence. The key is that humans do not use every option with equal probability.
If Netero has 100 possible defenses but, at one distance, chooses A 50% of the time, B 30%, and C 15%, most of the useful prediction problem is compressed into A, B, and C. If Meruem also attacks in a way that makes A more likely, the effective search space becomes even smaller.
The core idea is therefore not brute force. It is:
Limit the situation, detect behavioral bias, and use your own action to narrow the opponent’s response.
3. In markets, the “strong trader” is usually invisible
A one-on-one fight has one identifiable opponent. A market contains institutions, brokers, market makers, high-frequency traders, discretionary professionals, and retail traders at the same time. A normal public order book does not tell you exactly who submitted each order.
Therefore, “copy the best scalper” is a dangerous framing. What you can observe is not the person but the trace left in public data:
- How many ticks price moved after aggressive buying.
- Whether aggressive selling failed to push price down.
- Whether depleted asks were quickly replenished.
- Whether bids stepped upward.
- Whether trading continued after a new high.
- Whether the spread stayed tight or suddenly widened.
The market version of the question is not “Who is buying?” but:
Which side is actually moving price in this observable state?
4. Why a remora-style strategy may be possible
Several research findings support the broad idea of not racing for the first move, but trying to capture a remaining part of a persistent flow.
First, large orders are often split. Executing a large order at once can move the price against the trader, so a parent order may be divided into many child orders. A study using more than nine years of Tokyo Stock Exchange account-level data quantitatively supported the idea that order splitting helps create persistent order signs—buy orders tending to be followed by buys and sells by sells.[2]
Second, predictable child-order patterns can be anticipated. Research finds that more predictable large-order execution patterns are associated with stronger order-anticipation activity, higher execution costs, and delayed price impact as other traders gradually learn about the order.[6][7] This is related to “back-running”: using past order flow to infer and trade after a larger investor’s activity.
Third, the first speed race is far too fast for a human. A study of UK exchange messages found that typical latency-arbitrage races lasted about 5–10 microseconds, and more than 20% of trading volume occurred in races of some kind.[9] A person watching a screen cannot realistically compete for that first fragment.
A retail trader can only hope to study effects that last longer: order splitting, follow-on orders, and temporarily insufficient liquidity over seconds or minutes.
This does not prove that buying every “second wave” is profitable. Once a large order ends, some of its price impact may reverse, and the initial buyer may sell to late followers. The remora strategy is not based on the fact that price moved. It depends on evidence that the flow has not finished.
5. Four observable traces
5-1. Order Flow Imbalance (OFI)
OFI combines aggressive trades with additions and cancellations at the best bid and ask to measure the difference between buying and selling pressure. In a study of 50 US stocks, short-interval price changes were more robustly related to OFI than to raw trading volume, and the same imbalance moved prices more when market depth was low.[3]
In a tug-of-war analogy, OFI does not merely count how many people joined. It tracks which side the rope actually moved toward and whether people joined or left either side.
5-2. Queue imbalance
Let the quantity at the best bid be B and at the best ask be A:
QI = (B - A) / (B + A)
QI is near 1 when the visible best bid is relatively large and near -1 when the best ask is relatively large. Research on ten Nasdaq stocks found a statistically significant relationship between queue imbalance and the direction of the next mid-price move, with larger improvements for large-tick stocks.[4]
But visible depth can be cancelled. A still image of the book is not enough; additions, cancellations, and executions must also be observed.
5-3. Price response to orders
The most intuitive signal is not order size alone but how price responds to the order.
- Heavy aggressive buying without an upward move may mean a strong seller is absorbing demand.
- Aggressive selling without a downward move may mean buying interest is absorbing supply.
For a long continuation trade, “many buys” is weaker evidence than the market surviving a selling counterattack without breaking down.
5-4. Order-book resiliency
Resiliency describes how quickly the order book rebuilds after being depleted. One London Stock Exchange study found that the book did not always replenish, but when it did, the estimated half-life was around 20 seconds; the paper also reported evidence of copycat market-order submission.[5]
This is not a universal 20-second rule. Market, stock, and technology matter. The general lesson is to measure whether opposing liquidity returns during the brief period after the initial shock.
6. Entry hypothesis: capture reacceleration after survival, not the first burst
A simple long-side remora hypothesis has five stages:
- Shock: price, trade frequency, and positive OFI rise unusually fast.
- Wait: do not chase immediately; allow selling pressure to appear.
- Survival test: the pullback is shallow, bids replenish, and the origin of the move remains intact.
- Reacceleration: positive OFI returns, asks are removed, and the recent high is broken.
- Entry: participate only in the short continuation after reacceleration.
For example, price rises from 1,000 to 1,010, pulls back to 1,007, refuses to break below 1,007 despite selling, and then moves above 1,010 again. The fast trader receives the first ten points; the remora tries to take only part of what remains after confirmation.
This five-step pattern is not a published universal winning rule. The literature directly supports persistence in order signs, short-horizon information in OFI and queues, and the existence of order anticipation. Combining those findings into this exact entry rule is a testable strategy hypothesis.
7. Profit-taking should follow the life of the entry reason
The reason for entering is:
“A short-lived buying flow still exists, so follow-on buying may arrive soon.”
The exit should therefore depend not only on the amount of unrealized profit, but on whether that hypothesis remains alive. Four exit types can be combined.
7-1. Exit when the premise fails
Possible failure signals include:
- Price immediately falls back below the breakout.
- The pullback low breaks.
- OFI reverses toward selling.
- Aggressive buying continues but price no longer rises.
- The ask repeatedly replenishes at high speed.
- Bids disappear and the spread widens.
If the premise fails, exit even if the profit target has not been reached. If the fast runner you followed stops, there is no reason to continue alone simply because you planned a certain distance.
7-2. Time stop
The strategy expects continuation soon. If price does not make progress within the expected horizon, the lack of movement is itself information.
There is no universal “20 seconds for every stock.” Predictive signals decay at different rates, and dynamic trading research with transaction costs shows that optimal adjustment depends on how persistent or fast-decaying a signal is.[12] Measure how quickly winning setups normally begin to work, and use that empirical shelf life as the time stop.
7-3. Fixed target
A fixed target is not inherently bad, but arbitrary numbers such as “three ticks” or a universal 1:2 reward-to-risk ratio are weak. Collect the maximum favorable excursion (MFE) after entry and compare targets after spreads, fees, and slippage.
A simplified expectancy formula is:
Expectancy = win rate × average win − loss rate × average loss − trading costs
A distant target offers a larger payoff but a lower hit rate. Both must be considered together.
7-4. Trailing and partial exits
A trailing exit raises the stop as price advances, preserving some exposure to rare large moves. Under a specific diffusion framework, mathematical research shows that a sell limit combined with a trailing stop can be optimal.[11] This does not imply that “two ticks below” is universally optimal.
Partial exits may reduce outcome variability and psychological pressure, but they also sell part of the best trades too early and increase execution costs. There is no general proof that taking exactly half off is always superior. Full fixed exits, time stops, trailing exits, and scaling out must be compared on the same data.
8. A tighter stop is not automatically better
Stop-loss research shows that when prices are random walks, mechanical stop rules reduce expected returns, while they can add value when returns display positive serial dependence or momentum.[10]
Because the remora setup is explicitly a short-term continuation hypothesis, a premise-failure stop may fit it. But a stop narrower than normal market noise will be triggered repeatedly, leaving only spreads and fees.
It is useful to separate three roles:
- Initial stop: the price or order-flow condition that invalidates the setup.
- Time stop: the point at which expected continuation has failed to arrive.
- Profit-stage stop: a trailing level based on the latest pullback or flow reversal.
9. Human profit-taking is biased
A famous study of roughly 10,000 brokerage accounts found that investors tended to sell winners too early and hold losers too long—the disposition effect.[13]
The same bias can appear in scalping: a trader immediately takes a two-tick gain to avoid losing it, but waits for a two-tick loss to recover. Wins become small and losses become large.
Exits should therefore be linked to predefined reasons:
- The premise failed.
- The setup expired in time.
- A statistically chosen target was reached.
- A trailing condition was hit.
10. What research supports—and what it does not
Relatively well-supported findings
- Opponent modeling is a serious research approach for adapting to other agents.[1]
- Order splitting contributes to persistent same-direction order flow.[2]
- OFI is strongly related to short-horizon price changes.[3]
- Queue imbalance can contain information about the next price direction.[4]
- Other traders can anticipate predictable large-order execution.[6][7]
- The fastest market races are inaccessible to human reaction speed.[9]
- The usefulness of stops and trailing exits depends on the price process, signal persistence, and costs.[10][11][12]
Not generally proven
- That a shallow pullback and second breakout reliably produce net retail profits.
- That public book data identifies a particular expert scalper.
- That 20 seconds, three ticks, or a 50% partial exit is universally optimal.
- That prediction accuracy automatically becomes net profit.
- That a rule found in one regime remains valid after participants or market design change.
11. Minimum viable test design
Before using complex AI, record 100–300 examples of one precisely defined setup.
Context data
- Instrument, time, spread, and tick size.
- Size and duration of the initial move.
- Trade frequency and OFI or a simplified pressure measure.
- Pullback size.
- Price response to opposing market orders.
- Bid/ask additions, cancellations, and replenishment speed.
- Whether reacceleration and a new high occurred.
Post-entry data
- P&L after 1, 5, 10, 20, and 30 seconds.
- Maximum favorable excursion (MFE).
- Maximum adverse excursion (MAE).
- Time to peak.
- Time to OFI reversal.
- Actual fill, slippage, spread, and fees.
Compare exit policies
- Fixed target and fixed stop.
- Premise failure plus time stop.
- Partial fixed target plus trailing remainder.
- Full exit on order-flow reversal.
Split training and evaluation periods by date. Create rules on older data and test them on newer data without leaking the future. Repeat through walk-forward testing because market regimes change.
Always compare results after costs. A one-tick forecasting edge is not a strategy if one tick is lost to spread and slippage.
Conclusion
The main lesson from Netero versus Meruem is not to memorize every possible sequence.
Learn bias from visible behavior, narrow the situation, confirm the other side’s response, and enter only the short path that remains.
In a market, this does not mean finding a named “super trader.” It means reading the public traces left by strong order flow. Give the first move to faster participants, wait for the market to survive a counterattack, and join only if order pressure reaccelerates.
Profit-taking should then answer three questions:
Does the reason for following still exist? Did continuation arrive within its expected lifetime? Is enough movement left after all costs?
The shortest version is:
Do not try to beat the leader. Confirm that the flow created by the leader is still alive, ride it briefly, and leave as soon as the flow stops—regardless of an arbitrary distance target.
That is the most defensible research-based skeleton of a remora-style scalping strategy.
