Where the Money Sat in October
London, the third week of October 2026, and abrdn’s quarterly call has reached the slide that matters most to its retail army: customer cash. Siobhan Boylan, the group’s finance chief, puts the figure at £8.3 billion, roughly 7.5% of the assets held on the interactive investor platform. [1] Part of that money is waiting to be invested. Part of it has drifted into gilts because of where interest rates sit. Nobody on the line needs the term explained.
The platform is the growth engine, and the quarter was a strong one. Interactive investor drew £2.5 billion of net new money in the three months to September, 32% more than a year earlier. [1] Across the first nine months the intake reached £9.3 billion, up 58%. Customer accounts rose 8% to 533,000. Self-invested personal pensions — the tax-wrapped accounts UK savers use to hold funds directly — climbed 34% to 131,000. Platform assets reached £112 billion, 19% higher since January, and average daily retail trades ran at 28,000, 6% above the prior year.
The Adviser division supplied the counterweight. It shed £0.7 billion of net money in the quarter, unchanged from the second quarter and slightly worse than a year earlier. Adviser assets still rose 5% to £85 billion, but the gain came from markets rather than from clients. Gross flows into Wrap, the packaging service through which advisers place client money, reached £4.1 billion in the first nine months, the best level in five years. The company’s own assessment is that a return to positive net flows is taking longer than it had anticipated. [1] At group level, assets under management and administration stood at £579 billion at the end of September, up 4% since January, and again the increase came mainly from market movement rather than from new clients. The 2026 targets were restated unchanged: at least £300 million of adjusted operating profit and about £300 million of net capital generation. Inside institutional and retail wealth, money went to fixed income (plus £0.9 billion), real assets (plus £1.1 billion) and alternatives (plus £0.8 billion), while equities lost £1.9 billion and insurance partners bled £2.5 billion on the run-off of Standard Life’s legacy book. Then came the housekeeping: half of the Standard Life stake sold for about £430 million, a single-company concentration of nearly £1 billion cut back to a retained holding of just over 5%, and £210 million of additional Tier 1 debt — a subordinated bond that counts toward capital — earmarked for redemption at its first call date in December 2026. Debt reduction sits first in the queue for the proceeds.
Seventeen Thousand Transcripts Go In
Every quarter, calls like that one become documents. Zhiqiang Ma, Grace Bang, Chong Wang and Xiaomo Liu, researchers at S&P Global, assembled 17,025 earnings call transcripts from 485 companies in the S&P 500, spanning 2009 to 2019. Fifteen index members fell out for mergers, ticker changes or too short a run of available transcripts. The average company still contributed about 35 calls, and the raw material was enormous: the answer sections alone held 3.6 million sentences and 21.7 million tokens once stop words were stripped out, with a typical transcript carrying 2,145 sentences and 1,457 tokens.
The design rests on a judgement about human behaviour. Prepared remarks are scripted and rehearsed, especially when the news is bad; the question-and-answer exchange is not. So the model reads only the answers. Each sentence is turned into a vector by averaging and max-pooling pretrained 300-dimensional word embeddings — a list of numbers that positions a word by the company it keeps. An attention layer then assigns every sentence a weight, letting the classifier lean on the passages it finds informative and discount the rest. Transcripts are truncated or padded to 300 sentences, and anything under ten sentences is discarded.
Industry enters as its own learned vector, drawn from the eleven sector categories of the Global Industry Classification Standard. A feed-forward network — batch normalisation, dropout, ReLU activations, linear layers — compresses the combined text and sector features into a single bit: up or down. The target is the direction of the closing price on the trading day after the call. The test set is the five most recent transcripts from each company, 2,425 observations held out entirely from training.
Three baselines were built to be beaten. A 60-day mean-reversion rule — buy what has fallen below its own average, sell what has risen above it — scored 50.80% accuracy, with a Matthews correlation coefficient of 0.0202, a score where zero means no better than chance and one means a perfect match. XGBoost, a standard tree-based method, scored 50.89% (0.0013) on raw token counts and 51.25% (0.0154) on TFIDF weights. The attention model reached 52.45% and 0.0445, better by more than a percentage point on accuracy and roughly double on the correlation measure. Adding the prepared presentation text did not improve it, and a model fed the presentation alone did not improve either — a result that sits awkwardly against earlier work on volatility forecasting, where the scripted section carried more signal. The study records the discrepancy without settling it.
The sector split is where the numbers grow teeth. Information technology reached 56.8% accuracy, the best of the eleven categories; energy reached 48.5%, the worst, and below even odds. The reading offered for the spread is straightforward: technology prices react to bullish and bearish signals arriving from many directions, while energy prices track crude oil and macroeconomic forces more closely than they track any single company’s quarterly conversation.
Fifty-Two Percent Is a Whisper
Set the headline number against the difficulty of the task and it shrinks. A binary forecast of a share price one day forward, made from text alone, lands at 52.45% on 2,425 held-out calls. That is 2.45 percentage points above a coin flip and 1.65 points above a mechanical 60-day trading rule. The researchers are blunt about the terrain: “Undoubtedly, stock price movement prediction is a very challenging task.” What the model delivers is narrow by construction. It predicts direction, not size. It predicts the day after the call, at the close, and nothing beyond that horizon. A reader hoping to learn how far a stock might travel, or how violently it might swing, gets nothing here; volatility is a separate question the study does not attempt, and the earlier work it cites was aimed at exactly that different target.
The gap the researchers name themselves is the missing price history. Their proposed extension is to encode past price series with a recurrent neural network — a model built for sequences — and merge those vectors with the text features in the same forecast. That work is not done in this paper. Nor is the presentation puzzle resolved; the authors flag the contradiction with the volatility literature as a matter for future investigation. The dataset stops in 2019 and the study dates from 2020, so the market it describes is the one that existed before the pandemic repriced everything.
The limitation is the interesting part. A model that reads only language, with no access to prices, charts or order flow, extracts a consistent two-point edge from a document that companies prepare in advance and publish to the world. That edge is small, and its smallness is information: it says the words on a call carry a real but thin signal about the next session, and it locates the missing mass precisely where the researchers propose to look next — in the price history they have not yet wired into the machine.