Can a retail investor see machine-speed manipulation coming?

Between January 2023 and March 2025, on 15 of the 18 expiry days examined, one trading firm allegedly moved India's Bank Nifty index up in the morning and down in the afternoon. According to India's market regulator, SEBI, Jane Street bought large amounts of Bank Nifty stocks and futures before noon, often above the last traded price, while building options positions that would pay off if the index fell. On SEBI's example day, 17 January 2024, it sold after 11:49 a.m. (Kansal; SEBI, Interim Order). In July 2025, SEBI barred the firm from Indian markets and ordered it to deposit ₹4,843.57 crore, about $567 million at the time, while the case continued (Kansal). Jane Street paid the money into escrow, and the ban was lifted on 21 July 2025 (SEBI, "Update"). The same regulator had already found that 93 percent of individual traders in Indian equity derivatives lost money in fiscal years 2022 through 2024, with combined losses above 1.8 lakh crore (SEBI, "Updated SEBI Study"). SEBI does not attribute those losses to manipulation. Its latest study found that 87.7 percent still lost money in fiscal year 2026 (SEBI, Profitability). Jane Street disputes the allegations, and the order is interim: in October 2026 the firm was contesting it at India's Securities Appellate Tribunal, and SEBI had issued no final ruling ("Jane Street Accuses"). The case is an example of the problem this paper studies: a strategy that runs at machine speed, on public markets, against traders who have no way of seeing it.

High-frequency trading (HFT) uses computer programs to execute orders in microseconds or milliseconds and accounts for roughly half of U.S. equity trading volume, depending on how it is defined, yet public data does not show when it is happening (SEC Staff; Ibikunle et al.). It is run by proprietary trading firms such as Citadel Securities, Virtu Financial, Jane Street, and Hudson River Trading, which pay large fees to place servers inside exchange buildings and receive price data before regular investors do (Citadel Securities; Global Trading; IFLR; Duke Law & Technology Review). HFT strategies fall into three main groups: market making, which profits from the gap between buy and sell prices; latency arbitrage, which trades on price changes before slower exchanges update; and momentum ignition, a manipulative strategy that triggers a sharp price move, lets other traders follow it, then reverses (SignalPilot; Tse, Lin, and Vincent). Small traders appear to take the larger losses: in the S&P 500 futures market in 2010, aggressive HFTs earned $3.49 per contract from small traders compared with $1.92 from large fundamental traders (Baron, Brogaard, and Kirilenko). HFT also narrows bid-ask spreads, which lowers trading costs for retail investors (Hendershott, Jones, and Menkveld; SEC Staff), but Ibikunle et al. find that liquidity-demanding HFTs go with slower price discovery around earnings announcements (Ibikunle et al.).

Regulators and researchers have worked on this problem for more than a decade. The SEC built the Consolidated Audit Trail so that regulators can track every order in U.S. stocks (SEC, "Rule 613"), and enforcement actions have followed, from the SEC's first HFT manipulation case, against Athena Capital in 2014, to the first criminal spoofing conviction in 2015 (SEC, "SEC Charges"; U.S. Attorney's Office). Academic researchers have described the patterns manipulation leaves in the data and have trained machine learning models to find them (Tse, Lin, and Vincent; Golmohammadi, Zaiane, and Díaz; Cao et al.).

All of this work is built for someone other than the retail investor. The Consolidated Audit Trail is open only to regulators. Most academic models are trained on tick-level order book data or confidential trade records that cost thousands of dollars or cannot be bought at all. Enforcement arrives years after the trading: Athena's manipulation took place in 2009 and was charged in 2014 (SEC, "SEC Charges"). A retail investor deciding whether to buy a stock this afternoon has nothing that tells them whether its price is being pushed.

This paper asks two questions. First, what does existing research show about how HFT manipulation works and how much it costs retail investors? Second, can the detection methods that research describes be adapted to the minute-level price and volume data that a retail investor can get for free? Section 2 explains how legitimate HFT strategies make money and where manipulation begins. Section 3 reviews five manipulative strategies. Section 4 weighs the evidence on retail harm against the evidence on benefits. Section 5 reviews detection methods, from regulatory surveillance to machine learning. Section 6 presents the methods and results. Section 7 concludes. Beyond the existing reviews, the paper adds a strategy-by-strategy account of what free minute-level data can and cannot show, and a test of how a simple detector behaves on real data, with controls.

Most HFT activity is legal, and some of it helps other investors. In market making, a firm posts a buy order and a sell order at the same time, for example a bid at $100.00 and an offer at $100.05, and earns the five-cent gap plus a small rebate from the exchange when both sides fill (SignalPilot). The risk is adverse selection: a better-informed trader can hit the firm's quote just before the price moves against it. Brogaard, Hendershott, and Riordan studied NASDAQ data on HFT trades and found that HFTs, overall, trade in the direction of permanent price changes and against temporary ones, which means they help prices reach their correct level faster.

Latency arbitrage is harder to defend. When a price changes on one exchange, a firm with faster connections can trade on another exchange before that exchange's quote updates (SignalPilot). The profit comes from information other traders have not received yet. Budish, Cramton, and Shim argue that this is a flaw in market design. Because U.S. exchanges process orders one at a time in a continuous stream, a firm that is even a microsecond faster can always pick off stale quotes, so firms keep spending money on speed that does nothing for anyone else. Their proposed fix is to run auctions in short batches, so that speed differences smaller than the batch interval stop mattering. Aquilina, Budish, and O'Neill later measured these races directly in London Stock Exchange data: for FTSE 100 stocks they found about one race per minute per stock, with races making up about 20 percent of trading volume and costing traders roughly 0.5 basis points on average.

Ibikunle et al. show that the HFT population splits along this line. Using machine learning measures of HFT activity for U.S. stocks from 2010 to 2023, they find that HFTs who supply liquidity are associated with better price discovery after earnings announcements, while HFTs who demand liquidity are associated with worse price discovery.

Manipulation is a narrower category. Putniņš defines market manipulation as trading or order placement meant to move a price away from where supply and demand would put it, so that the manipulator can profit from the distortion. His categories include trade-based manipulation, which uses real trades, and order-based manipulation, which uses orders the trader never intends to fill. Both types appear in HFT. What makes HFT manipulation different from older forms, such as pump-and-dump schemes, is its scale and speed: a single algorithm can repeat the same pattern hundreds of times a day across many stocks, and each episode can be over in seconds. The five strategies in Section 3 rest on different kinds of evidence: Coscia's conviction was adjudicated, Athena settled without admitting or denying the findings, the Jane Street order is a pending allegation, and momentum ignition and quote stuffing are patterns measured in data, not proven conduct.

The most direct measurement of who loses to HFT comes from Baron, Brogaard, and Kirilenko, who used confidential regulatory data that tags the trading account on both sides of every trade in the E-mini S&P 500 futures contract. HFTs earned more than $23 million in gross trading profits in that contract in August 2010 alone. Aggressive HFTs, the ones that take liquidity rather than provide it, earned $3.49 per contract when trading against small traders and $1.92 per contract against large fundamental traders. Small traders therefore paid about 82 percent more per contract than the larger traders. These are short-term profits, which the authors read as a trading cost for the other side, and they come from a 2012 working paper; the published version of the study uses Swedish stock data instead and does not report them (Baron et al.).

The harm is not spread evenly across retail investors. A Congressional Research Service report, citing a Canadian study, noted that retail investors who day trade may be more affected by HFT, while buy-and-hold investors who trade rarely are less likely to be affected (Shorter and Miller). That Canadian study, by Malinova, Park, and Riordan, is mixed on retail traders: after a 2012 fee on messages cut message traffic by about 30 percent, institutions' costs on market orders rose, retail market-order costs did not change, and retail returns on limit orders fell. The strategies in Section 3 explain why. A momentum ignition event costs something only to the trader who buys during the artificial rise, and a person who buys an index fund once a month is rarely that trader. An active trader following short-term price moves is exactly the trader momentum ignition is designed to attract.

Speed differences add to the cost. Jian, Yin, and Li built an agent-based simulation of the Chinese market, using order book data to set up the simulated market, and found that when the gap in communication delay between retail traders and HFTs exceeded 50 milliseconds, the chance that a retail order would be intercepted through latency arbitrage rose to over 82 percent. This is a simulation, not a measurement of real losses, and the 50-millisecond threshold is built into the model, but it shows how quickly the disadvantage grows once the speed gap passes a threshold. The one large measurement from real trading data puts the cost per trade at about 0.5 basis points, or 5 cents on a $1,000 trade (Aquilina, Budish, and O'Neill). That is about the same size as the futures figure above: $3.49 on a contract worth about $50,000 is roughly 0.7 basis points.

HFT can also make market-wide stress worse. Kirilenko et al. studied the Flash Crash of May 6, 2010, when U.S. stock prices collapsed and recovered within minutes. They found that HFTs did not trigger the crash but that their trading, in response to an unusually large sell order, made the volatility worse.

Other parts of market structure also shape what retail investors pay. Schwarz et al. placed 85,000 simultaneous market orders through six accounts at five U.S. brokers and found that average round-trip costs ranged from 0.07 to 0.46 percent, because the wholesalers that execute retail orders gave different brokers different prices for the same trades. In U.S. options, retail traders now account for over 60 percent of volume, favor cheap weekly contracts with an average bid-ask spread of 12.6 percent, and lose money on average (Bryzgalova, Pavlova, and Sikorskaya).

Against this evidence, HFT competition has lowered the direct cost of trading. Algorithmic trading, including HFT market making, has narrowed bid-ask spreads (Hendershott, Jones, and Menkveld; SEC Staff), and liquidity-supplying HFTs make prices more accurate (Brogaard, Hendershott, and Riordan; Ibikunle et al.). The two bodies of evidence may be compatible, but the evidence on retail traders is mixed, as the Canadian study above shows. Every retail trade may be slightly cheaper because of HFT market making, while the retail traders who trade most often, and at the worst moments, lose more to aggressive and manipulative strategies. Averages across all investors hide the second effect. The problem this paper addresses is the second group.

Regulators have the best data. The SEC adopted Rule 613 in 2012 and approved the plan for the Consolidated Audit Trail in 2016, creating a central database that lets regulators track all activity in U.S. listed stocks and options, including which firm placed each order (SEC, "Rule 613"). The system is now being scaled back: a federal appeals court struck down its funding order in July 2025 (American Securities Association v. SEC), the SEC has stopped CAT reporting of individual customers' personal information, and in August 2026 SEC Chairman Paul Atkins asked staff to draft a rule that would rescind Rule 613, a change he expected to take until late 2027 (Atkins). With that data, a regulator can see a spoofing pattern directly: the same firm placing and cancelling large orders just before trading on the other side. The limits are access and timing. The data is available only to regulators, and enforcement takes years. Athena's 2009 trading was charged in 2014, Coscia's 2011 trading ended in a 2015 conviction, and SEBI's order against Jane Street came almost a year after the last of the 18 expiry days it examined, with no final ruling by October 2026 (SEC, "SEC Charges"; U.S. Attorney's Office; SEBI, Interim Order; "Jane Street Accuses"). Enforcement punishes manipulation after the fact and gives no warning to the investor who is trading against it.

The simplest detection methods write the known pattern as a set of rules. Tse, Lin, and Vincent describe momentum ignition as a sequence: a volume spike with little price change, a sharp price move with more volume, and a reversal. A rule-based detector flags any window where all three conditions are met. Rules are easy to explain, which matters for a non-expert user, but they only find the patterns someone has already described, and a manipulator who changes one step of the sequence can avoid them.

Supervised models learn from examples labeled as manipulated or not manipulated. Golmohammadi, Zaiane, and Díaz used stocks named in SEC manipulation cases from 2003 as their positive examples and tested eight algorithms, including random forests, neural networks, and support vector machines, on hourly trading data from 64 companies. Naive Bayes performed best, catching 89 percent of manipulated observations with 83 percent specificity. The weakness is the labels. Confirmed manipulation cases are rare, they come from enforcement actions years after the trading, and they include only the cases regulators caught. A model trained on them learns what caught manipulation looks like.

Unsupervised models do not need labels. They learn what normal trading looks like and flag what does not fit. Cao et al. built an adaptive hidden Markov model with extra "anomaly states" for price manipulation and tested it on tick-level data with simulated manipulation cases, where it outperformed the benchmark methods it was compared with. Isolation Forest, introduced by Liu, Ting, and Zhou, takes a different approach. It builds many random decision trees that split the data on random features at random values. Unusual points get separated from the rest in fewer splits than normal points, so the average number of splits needed to isolate a point becomes its anomaly score. The method is fast, needs no labels, and works with a small number of features, which makes it a good fit for data with few columns.

Recent studies have tested unsupervised models on data closer to what a retail investor has. Núñez Delafuente, Astudillo, and Díaz applied Isolation Forests to hourly price, return, volume, and trade-count features for eight stocks named in SEC cases. A single forest found about two-thirds of the manipulated hours, and an ensemble of forests found up to 89 percent, but only 2 to 6 percent of the hours they flagged were actually manipulated. In a 2026 preprint, Chen and Hybinette ran minute-level detectors on public Bank Nifty options data covering 209 trading days in 2024, 10 of which SEBI identified in the Jane Street case. An autoencoder flagged all 10 and an Isolation Forest flagged 8, but about three of every four flagged days had no known manipulation. Legitimate trading can also produce similar shapes: dealers hedging same-day options can strengthen intraday reversals and momentum on their own (Dim, Eraker, and Vilkov).

The study closest to the retail problem is Ibikunle et al. They trained machine learning models on a proprietary dataset in which HFT activity was identified, then applied those models to public intraday data to estimate HFT activity for U.S. stocks from 2010 to 2023. Their measures performed better than older proxies and responded correctly to known events such as exchange speed changes. Their goal was measurement, not manipulation detection, but they show that public data carries enough information to estimate HFT activity when a model is trained properly.