Intraday alternative data has long been the preserve of high-frequency and systematic desks, where speed provides an edge. But ask practitioners what decision it actually changes for a discretionary book, and the answer is rarely about generating a new trade, it is more about testing the one already on. Sizing, conviction and the risk sitting on the book are where intraday signals earn their place – not in producing fresh alpha, but in interrogating positions a desk already holds.
A panel at the recent A-Team/Eagle Alpha Alternative Data Conference in London brought together Mark Fleming Williams, Head of Data Sourcing at CFM, who moderated; Davide Alfano, Managing Director at Kaleidoscope Capital; Mario Dell’Era, Senior Quantitative Market Risk Manager at EnBW; Joe Gits, Managing Director of North America at BridgeWise and CEO of Context Analytics; and Ovie Koloko, Chief Product Officer at Parameta Solutions, the data and analytics division of TP ICAP.
For a discretionary investor, the argument ran, intraday data is less a trigger than a check on whether current positioning still holds: a reason to adjust ahead of a quarterly release, or to read sentiment before running a block trade. That holds for systematic strategies on medium-term trends too, where intraday inputs prompt small adjustments rather than a wholesale rethink.
The clearest version of the point came from the risk side: intraday data is used not to build strategies but to stress-test them, running existing positions against such inputs to find where they give way. On that view signal decay barely matters, because the data captures current stress rather than a tradable edge with a shelf life.
Cadence Sets the Clock
One idea ran through the technical discussion: “intraday” describes a use case, not a property of the data. What governs how a signal behaves is the cadence of the underlying source – the rhythm at which it is generated – not the frequency at which it is sampled.
Social media was the worked example. Because the conversation moves in seconds, the signal suits short-horizon aggregation, with measurable return deviations over the following 15 to 20 minutes; widen the window and the same feed yields a slower, multi-day read. A lower-cadence source such as supply-chain data cannot support that fast a horizon at all, and sits more naturally in security selection. Noise falls under that principle too: high-frequency social data is heavily polluted by spam and bots, but the answer is account-level scoring and filtering, not abandoning the source. The tooling for measuring decay – information coefficients, heat maps, monotonicity checks – is well established. The failure mode lies elsewhere.The Real Constraint Is Use-Case Discipline
That failure mode is mismatch. Effectiveness, the discussion suggested, is rarely the real problem – credit card and unstructured social data routinely test well. Trouble starts when a dataset built for one time horizon is pressed into another. Aggregation bias, timing cuts and the temporality of the data are where daily datasets quietly fail when stretched to intraday use. Market microstructure data – pricing, liquidity, order-flow proxies – travels well from back test into production, because it is grounded in observable market activity rather than inferred behaviour.
Signal decay, on this reading, is less about crowding than about the data shifting underneath the user. One line of argument weighted schema and structural change above competitive erosion: a signal that stops working shows up in performance, but a dataset whose construction quietly changes can degrade a whole process without immediate feedback. Tracking how the data is built matters more than tracking whether a given signal still pays. Pricing fell along the same lines, with the notion that price signals quality rejected in favour of a model anchored to what a client intends to do with the data.
Where decay does bite, governance was framed as a shared responsibility rather than any one team’s. Research chooses and maintains the signals, data engineering handles collection, storage and quality, and PMs and traders own execution – the now familiar trust-but-verify model. Kill-switch frameworks, regime-change detection and an explainability layer linking a signal to a macroeconomic rationale were named as the apparatus that has to surround an intraday signal before it touches capital. The starkest danger was silos: a strong data team with a disengaged PM, or a PM with appetite but no structure behind them.
The Judgement Layer AI Cannot Reach
As often happens, the discussion turned to AI – and the same use-case discipline reappeared. On the vendor side, AI was cast as a product layer, not a research engine: machine learning has long underpinned the back-tested feeds, which must stay reproducible and auditable, while generative tools repackage those feeds into summaries aimed at PMs rather than quants. The back test itself stays off-limits to generative AI, as it has always demanded reproducibility.
On the buy side, the same question – what is the right use, and who exercises the judgement – carried into how desks are absorbing AI. The framing was augmentation, not replacement: AI does not take the job; someone using AI does. Productivity gains for senior and junior developers alike were acknowledged, but the skill threshold was seen to rise rather than fall – fewer people, more qualified, with code-review tooling and humans kept in the loop. The judgement that decides whether a signal should be trusted, scaled or retired is the same institutional knowledge, built over decades, that decides whether a dataset fits its use in the first place. That is what does not transfer to a model.
Which returns to where it started. Intraday alternative data does change decisions, but mostly defensive ones, and its value is bounded by the discipline of matching data to use case. The question left open is whether that judgement layer, the part no one was willing to automate, scales as fast as the data does.
Subscribe to our newsletter
