We don’t write press releases. We write what we actually think about where markets, AI, and private capital are heading — and what it means for those operating in them.
Most systematic programmes let you assume the money was real. Ours was not, and we say so at the top of the page rather than the bottom. Here is what a live-market paper account actually establishes, what it cannot establish at any size, and the worst trade we have ever printed — including why we published it.
Read article →Algorithmic strategies now account for over 70% of U.S. equity volume. What started as a hedge fund arms race has quietly become the infrastructure of modern markets. The question is no longer whether algorithms belong in a portfolio — it's whether you can afford to be on the wrong side of the machines that run it.
Read article →The firms winning with AI aren't using it as a faster Google. They've rebuilt their workflows around it. The distinction matters more than most people realize — especially in private equity, where the bottleneck has always been human bandwidth.
Read article →A senior associate at a top-decile PE fund will spend less time modeling and more time deciding. The firms building that infrastructure now aren't just getting efficient — they're building a structural advantage that compounds every deal cycle.
Read article →Most systematic trading programmes do not tell you how they were funded. They show a curve, a Sharpe ratio and a trade count, and they let you assume the money was real. Sometimes it was. Often it was not — and when it was not, the disclosure tends to appear in six-point type at the bottom of a page nobody scrolls to.
We do it the other way around. Every figure Obsidian Quant publishes was produced by a live-market paper account funded at $100,000 on 17 February 2026 — real market data, real prices, broker-simulated fills. No client capital has been deployed to date. That sentence sits near the top of our performance page, not buried at the bottom of it.
This piece is about why that distinction matters more than most people think, what a paper record does and does not establish, and what we believe you should ask of any systematic programme — ours included — before you deploy real money against it.
It is not a backtest. That difference is the entire point, and it is routinely blurred.
A backtest runs a strategy over history the researcher has already seen. Every choice about which signals to use, which parameters to set and which instruments to trade is made with the answer available. Even careful researchers leak information backwards; the ones who are not careful produce curves that have never survived contact with a live market and never will.
A live-market paper account runs forward, in real time, against prices that had not happened yet when the decision was made. The data is real. The prices are real. The clock is real. What is simulated is the fill: the broker tells you what would have happened to your order rather than actually working it in the book.
That last point is worth dwelling on, because it is where most systematic programmes quietly fail. A strategy that is sound on paper and an operation that survives six months of unattended running are different achievements. The second one is harder.
Here is the honest ledger, and it is longer than the one above.
A paper record is evidence about the decision logic. It is not evidence about execution. Anyone who conflates the two is selling you something.
On 19 August 2026 the system closed a single position at a realised loss of $17,678. It is on our performance page, and it will stay there.
The cause was not a market call. An order was placed while the exchange was closed. The system read the absence of a working order as a completed trade, cleared its own record of it, and allowed the same order to be placed again. The two filled together at the next open as one oversized position, which was closed the same day at that loss.
That is a defect in order handling, not in how positions are selected. The distinction matters — but it does not make the loss smaller, and we would rather state both facts than lean on the flattering half.
The fix is three refusals: an order is refused whenever the exchange is closed, refused again if an order is already working in the same security, and refused again if the resulting position would exceed its intended size. Each check fails to the safe side, none of them can block an exit, and a regression suite reproduces the original conditions to confirm the sequence cannot run again. No client account was affected, because there were no client accounts.
Because the alternative is worse. A performance page with no bad days on it tells a sophisticated reader one of two things: the record is too short to have had one, or the bad days have been removed. Neither builds confidence.
There is also a specific reason this particular loss belongs in public. Our largest equity drawdown — 17.1%, reached in early June 2026 — was also an order-handling defect, in protective stop placement, identified and corrected. Between that correction and 19 August the equity-based figure was 4.1%.
That cuts both ways, and we would rather you saw both edges. The favourable reading is that our worst two episodes were engineering failures rather than the strategy being wrong about the market. The unfavourable reading is that a system is only as good as the code that executes it, and ours has shipped two order-handling defects in six months. Both readings are fair. The second is the reason we now publish the fix alongside the failure.
Because a record this short cannot support one. Extrapolating a few months to a yearly figure produces a number that is arithmetically correct and analytically meaningless — and a sophisticated reader who sees it will discount everything around it, including the parts that were honest.
Returns on our performance page are presented cumulatively over the stated period and are not annualised, consistent with the convention that performance for periods of less than one year should not be extrapolated. Portfolio equity is shown normalised to a $1,000,000 starting portfolio using the same live return profile, which is a presentation choice, not a claim about deployed capital.
If you take one thing from this piece, take the list rather than the conclusion.
The last question is the one we would put first. Obsidian Quant is licensed technology: capital stays in an account the client owns and controls, we do not take custody, and we cannot move money out of it. That structure is not a performance claim — but it is the one thing on this page that does not depend on trusting our numbers.
A paper record proves less than most programmes imply and more than sceptics allow. It is worth exactly what it is: months of forward-dated evidence about how a system decides, published with its defects attached, waiting on the only test that settles the question — real money, at size, in a live book.
Every figure discussed here, including the drawdown detail and the full incident write-up, is published on the Obsidian Quant performance page. Past performance of the system is not indicative of future results. Trading involves substantial risk of loss.
See the performance page →In 2005, algorithmic trading accounted for roughly 25% of U.S. equity volume. By 2010, it had crossed 50%. Today, conservative estimates put it above 70% — with some asset classes, particularly futures and options, running closer to 80–90% automated. This is not a trend. It is the new baseline.
And yet, the conversation in most private wealth and family office circles still frames algorithmic strategies as exotic — something reserved for Renaissance Technologies and Two Sigma, not for the individual allocating $500,000 or the family office managing $50 million. That framing is outdated, and it's costing people money.
The first wave of algorithmic trading was purely about speed. High-frequency traders colocated servers next to exchange matching engines, shaved microseconds off execution, and extracted basis points at scale. That game still exists, but it's a closed loop — the arms race for sub-millisecond advantages is now dominated by firms spending tens of millions on custom FPGA hardware and dedicated fiber optic routes.
The second wave — the one that matters for everyone else — is about signal intelligence. Not speed, but the ability to process more information, more consistently, with fewer cognitive errors, than a human analyst making discretionary decisions under pressure. This is where the real edge lives for non-HFT systematic strategies.
"The human brain is not built for markets. It's built to survive. Those are different optimization functions."
Discretionary trading relies on pattern recognition, intuition, and experience — all of which are genuinely valuable, but all of which are also subject to loss aversion, recency bias, overconfidence, and the simple fact that human beings need to sleep. A systematic strategy doesn't have a bad week because of a difficult personal situation. It doesn't average down on a losing position because it's emotionally invested in the original thesis. It executes the same logic — consistently, repeatedly, at scale — regardless of conditions.
The most common misconception is that algorithmic trading means "following moving averages and RSI." That's 1990s quant finance. Modern systematic strategies layer multiple independent signal sources — price action, volume dynamics, market microstructure, regulatory intelligence, options market positioning — and look for confluence: the condition where multiple independent data streams agree on the same directional conclusion simultaneously.
The logic is elegant: any single indicator can generate false positives. A rising RSI can mean momentum or can mean an overextended move about to reverse. But when five independent signals converge on the same conclusion, the probability of a false positive drops dramatically. It's the difference between one witness and five independent witnesses telling the same story.
The addition of real-time regulatory intelligence is underappreciated. SEC EDGAR processes thousands of filings daily — Form 4 insider transactions, 8-K material event disclosures, Schedule 13D ownership changes. The market is inefficient at processing this data quickly. A system that can parse an 8-K the moment it hits EDGAR and extract a directional signal has a genuine edge window — typically 15 to 90 minutes — before the information is priced in by the broader market.
The weakest point of most algorithmic strategies is model decay. Markets evolve. Regime changes — from trending to mean-reverting environments, from low to high volatility, from bull to bear — can render a strategy that worked for three years suddenly ineffective. The traditional response is manual re-optimization by a quant team, which is expensive, slow, and often reactive rather than proactive.
The more sophisticated response is to build the re-optimization process into the system itself. An AI critique layer that reviews every closed trade — identifying which signals contributed to wins, which contributed to losses, and what adjustments to the decision logic would have improved outcomes — and proposes targeted rule modifications between trading cycles. Not autonomous rewriting, but a structured feedback loop that surfaces insights a human analyst would take weeks to identify.
This is the concept of recursive self-improvement applied to trading: a system that learns from its own history in a structured, auditable way, rather than repeating the same errors indefinitely.
For decades, the infrastructure required to run a serious systematic strategy — co-location, data feeds, execution management systems, risk frameworks — cost millions of dollars annually. It was genuinely inaccessible to anyone outside of institutional finance.
That has changed. Cloud computing, institutional-grade API-connected brokerages, and AI-powered development tools have compressed the infrastructure cost by orders of magnitude. The question is no longer whether serious systematic strategies can be deployed outside of a hedge fund structure. They can. The question is whether you're accessing that capability — or leaving it to others who are.
The market doesn't care whether your capital is deployed algorithmically or discretionarily. It only cares about the quality of your decisions at the moment you make them. Systematic strategies are not a guarantee of performance. But they are a structural approach to making better decisions, more consistently, at scale — and in a market where 70%+ of volume is already algorithmic, that matters.
Obsidian Quant licenses its algorithmic trading technology to qualified individuals and institutions. Your capital stays in your own account — we never custody it.
Explore Licensing →Most organizations are using AI wrong. Not because they're using the wrong models or the wrong vendors — but because they're treating AI as a productivity tool rather than as infrastructure. The firms that will look back on this decade as transformational are the ones that figured out the difference early.
A productivity tool is something you reach for when you have a specific task. You open it, use it, close it. A spreadsheet is a productivity tool. A calculator is a productivity tool. If you're using AI the same way — asking it to summarize a document, draft an email, or explain a concept, then moving on — you are capturing approximately 5% of the available value.
An operating system is something your entire workflow runs on top of. You don't use it for a task. It enables every task. The difference in leverage is not incremental. It's a different order of magnitude.
Private equity and growth equity due diligence is, at its core, an information processing problem. A typical deal process generates hundreds of pages of documents — CIM, financial model, management presentation, quality of earnings, legal data room — that need to be synthesized into a coherent investment thesis, risk assessment, and decision framework under significant time pressure.
The firms that close the best deals are not necessarily the ones with the best judgment. They are the ones with the best judgment and the fastest information processing. An investment team that can synthesize a CIM into an initial thesis in two hours rather than two days has a materially different decision advantage — they can run more deals, go deeper faster, and walk into an IC meeting better prepared than the competition.
"Speed doesn't just save time. In competitive deal processes, it changes which deals you can pursue."
This is not a hypothetical efficiency gain. An AI system trained on the structure of investment committee memos, LBO modeling conventions, risk frameworks, and deal dynamics can process a 150-page CIM and produce a full IC memo draft — with LBO model outputs, risk register, data gap analysis, and executive summary — in a fraction of the time a senior associate would spend on the same task.
Here's the distinction that matters: using ChatGPT to summarize a document is a tool interaction. Building a workflow where your deal origination pipeline feeds documents directly into an AI system that extracts the deal parameters, runs the financial model, flags the risk factors, generates the memo, and routes it to the IC — that is an operating system.
The difference is that the first interaction helps you one time with one document. The second interaction scales. Every deal that enters the pipeline benefits from the same system. The marginal cost of processing deal number 50 is essentially identical to deal number 1. You are compressing the economics of a 10-person deal team into a 2-person deal team — without sacrificing output quality.
The most common objection is that AI-generated analysis will miss the nuance that experienced deal professionals bring to a process. This objection misunderstands what AI systems are being asked to do. A well-designed system doesn't replace judgment. It handles the information structuring, the model mechanics, and the memo scaffolding — so that human judgment can be applied to the questions that actually require it.
What requires human judgment: is this management team credible? Is this market narrative defensible? Does the exit path make sense given current buyer appetite? Is the business actually performing the way the numbers suggest?
What does not require human judgment: extracting EBITDA figures from page 47 of a CIM. Building an LBO model from those figures. Formatting the executive summary. Organizing the diligence checklist. Tracking which data room documents have been received.
The firms building AI-native deal workflows are not automating judgment. They are automating the scaffolding around judgment — freeing their people to do more of the work that actually requires them.
Technology adoption curves in private equity are slow — slower than in most industries, because the asset class is relationship-driven and conservative by nature. But they're not infinitely slow. The firms building these capabilities now are establishing a structural advantage that will compound across every deal cycle. In three years, AI-native deal infrastructure will not be a differentiator. It will be a baseline expectation, the same way Excel models and Bloomberg terminals eventually became table stakes.
The question is not whether to build on AI infrastructure. It's whether you do it now, while it's still an advantage — or later, when it's a catch-up exercise.
DEALITHIC is Barenberg Capital's AI-powered deal engine — built for PE and growth equity teams who want to move faster without sacrificing rigor.
Explore DEALITHIC →Private equity has always been a talent-intensive business. The model — hire the best analysts from investment banking, work them relentlessly, filter for the ones who develop deal judgment, promote slowly — is as old as the industry. It produced extraordinary results for decades. It is also, quietly, becoming a structural liability.
The problem is not that the people are bad. The problem is the ratio of what they spend their time on to what they are actually worth. A third-year associate at a top-quartile fund commands a $350,000 total compensation package and spends approximately 40% of their time on tasks that, within five years, will be performed entirely by AI systems. The economics of that arrangement are going to shift — dramatically, and faster than most firms expect.
Break down the workflow of a deal team from CIM receipt to IC meeting and you find that the hours are heavily concentrated in a handful of tasks: initial CIM analysis and thesis formation, financial model construction, memo drafting, data room organization, diligence tracking, and management preparation. Of these, the first two have the highest human-judgment content. The rest are largely mechanical.
Memo drafting — the act of translating a financial model and a set of diligence findings into a structured investment committee document — is a template-driven process. It has a defined structure, a defined set of sections, and a defined logic for how information flows between them. A senior associate does it better than a junior one because they have more pattern recognition from having done it many times. An AI system trained on thousands of IC memos can develop that pattern recognition faster, apply it more consistently, and produce a first draft in minutes rather than days.
"The bottleneck in private equity has never been judgment. It's always been bandwidth. AI eliminates the bandwidth problem."
This is not a marginal productivity improvement. Cutting 40% of a senior associate's time on mechanical tasks doesn't mean you need 40% fewer associates. It means each associate can cover 40% more deals — which at a fund actively sourcing 200+ opportunities per year to close 4–6, that capacity expansion is the difference between passing on something great because you didn't have time to diligence it properly and getting there first.
The initial phase of AI adoption in deal teams is already underway. It looks like this: a junior analyst uses an AI tool to produce a first-pass summary of a CIM. A senior associate cleans it up and adds their own analysis. The IC memo takes three days instead of six. The financial model takes four hours instead of twelve. The team can run two processes simultaneously instead of one.
This is the productivity tool phase — real value, but not the structural shift. The structural shift happens when the workflow is redesigned around the AI capability rather than bolted on top of an existing workflow. When the CIM comes in and is automatically ingested, the deal parameters are extracted and fed directly into the financial model, the model outputs flow into the memo template, the memo template populates with deal-specific language, and the team's first interaction with the deal is reviewing and refining a nearly complete package — not starting from a blank spreadsheet.
The deal team of 2028 is not smaller necessarily — though some firms will go that direction. It is differently structured. The ratio of senior judgment to junior execution shifts dramatically. You need fewer people doing the mechanical work and more people with the experience to ask the right questions of the output — to identify when the model is making an assumption that doesn't hold, when the thesis has a fatal flaw that the AI missed because it was embedded in a footnote on page 83, when management's narrative doesn't reconcile with the customer concentration data on the data tape.
Those are not things that AI systems will do well by 2028. They require exactly the kind of pattern-matching, skepticism, and domain experience that takes years to develop. The professionals who develop those skills and layer them on top of AI-native workflow tools will be the most valuable people in the industry. The ones who resist the tools, or who allow their skills to atrophy because the tools are doing too much, will find themselves on the wrong side of a widening capability gap.
The conventional wisdom is that AI will primarily benefit large institutions with the resources to build proprietary tools. The opposite is closer to the truth. A $100 million fund with a three-person deal team, using AI-native deal infrastructure, can punch significantly above its weight class — sourcing, diligencing, and executing on deals that would have been impossible to process with a three-person team five years ago.
The technology doesn't care about the size of the AUM. It delivers the same leverage to the two-person emerging manager as it does to the billion-dollar platform. What it rewards is the willingness to redesign the workflow rather than layer the tool on top of the old one. The funds that figure that out — particularly at the smaller end of the market where each incremental deal capacity unit matters most — will compound that structural advantage deal by deal.
2028 is not far away. The teams building this infrastructure today are not future-proofing. They're winning now.
Barenberg Capital's advisory practice works with PE and growth equity teams navigating transactions where speed and rigor both matter.
Speak with Our Team →