EXAONE Forecast for Finance: A Linear Foundation Model Designed Apart for Finance

I read EXAONE Forecast for Finance, published on arXiv as arXiv:2609.04239. The central claim is that a financial time series foundation model called EXAONE Finance can rank first across point-forecast accuracy, cross-sectional asset ranking, and portfolio profitability on the FinVerse benchmark when it combines an attention-free linear backbone with causal temporal mixing, group-aware variate mixing, masked context augmentation, and pretraining across six market sleeves. The original paper is linked under the title.

This post traces that claim at walking pace. The first part asks why financial forecasting needs its own design, where self-attention backbones stall on finance panels, and what the attention-free choice changes. The middle part opens the three devices one by one, causal convolution for time, group-aware pooling MLP for variates, and masked context augmentation for missing spans. The later part reads the six-sleeve pretraining data and the three-tier FinVerse evaluation, weighs the constraints of missingness and length and channels in practice, and lists the questions that remain. Numbers are summit photos and process is the route map. This post spends most of its time on the route map.

Why financial forecasting needs its own design

The phrase time series forecasting is broad. Power demand is a time series, traffic flow is a time series, and factory sensors are time series. Financial time series have a different texture. Prices move through a tangle of expectations and fear and liquidity. Yesterday's pattern carries only a weak promise to repeat tomorrow. When the regime shifts, the whole distribution shakes. One rate decision, one earnings line, or one geopolitical event can change the character of a curve. So models that work well in general domains often hesitate in front of finance. The shape of the mean may look similar while the shape of the risk differs.

Second, finance is not a single curve but a panel. Hundreds of names trade at once, many markets overlap, and indicators press on each other in layers. Even inside equities, single-name moves sit inside sector moves, and sector moves sit inside the breathing of the whole market. Exchange rates and commodities and bond yields and macro indicators pull and push on each other. Crypto assets carry yet another texture of volatility. Forecasting means reading that tangle. It is not the smooth extrapolation of one curve but the reading of pressure that many curves place on each other. Where a general model grows strong on the long breathing of one or two channels, a finance model must grow strong on the short and noisy chorus of hundreds of channels.

Third, evaluation in finance never ends with one score. Closeness to tomorrow's close matters, but earlier questions press harder. Did the model rank risers above fallers correctly when names are set against each other, and does that ranking earn money when it is turned into a real portfolio. A model with good point forecasts but shaky rankings has little use on a desk. A model with good rankings that cannot survive turnover and cost has little use in an account. So financial forecasting research watches accuracy and ranking and profitability together. That EXAONE Finance aims at all three tiers at once catches the eye. It declares an intent to join curve fitting and rank building and money making inside one model.

Practitioners will nod at this point. Anyone who has touched finance data remembers the moment when compute burst on a long and wide panel. Extend the length and memory bursts. Extend the channels and time bursts. So teams compromise by cutting length or thinning channels. Yet the cut length may hold the clue to a regime turn, and the thinned channels may hide sector pressure. Every compromise removes something the model should have seen. Designing apart means reducing that compromise. It means building a body that can ingest financial length and width as they are. The rest of this post follows the shape of that body.

Where quadratic self-attention stalls on finance panels

The backbone of most existing time series foundation models is self-attention. Self-attention is a strong device. It connects two distant time steps directly and reads the tangle across channels with flexibility. In general domains that flexibility shines. When length is moderate and channels are few, attention weights read crisply. There is even pleasure in reading which step attended to which. So many studies built on top of the attention backbone. A winning formula is hard to replace.

The trouble is that cost grows quadratically with sequence length and variate count. Double the length and compute roughly quadruples. Double the channels and it quadruples again. On finance panels where length and channels grow together, that multiplication turns steep. Years of daily bars for hundreds of names already stretch the matrix. Stepping down to minutes stretches it further. Cut the length in half and compute falls to a quarter, but the breathing of regimes is cut with it. Thin the channels and compute falls, but the chorus of sectors thins with it. Researchers must weigh length against width on every run. A large model and a widely seeing model come apart at this point.

Look one step closer and quadratic cost is not only a story about slowness. It changes the research question itself. When compute is dear, teams avoid long-context experiments. When channels are many, they postpone cross-name experiments. Promising questions vanish one by one before the wall of compute. What would change if long macro flows and short single-name jerks were fed together. Such a question cannot be answered by split feeding, yet joint feeding bursts the budget. So the backbone choice is not a speed choice but a question choice. Which questions a model can ask is decided at the backbone.

Another layer is financial noise. Self-attention is as candid toward noise as it is flexible. On noisy panels, attention weights wobble from place to place. One spike pulls the weights, then the next stretch pulls them back. On the smooth curves of general domains, with little noise, that wobble stays small. On the bumpy curves of finance, it grows. Flexibility turns into sensitivity at this point. Taming that sensitivity calls for many surrounding devices, more normalization and masking and augmentation. More devices mean heavier models and harder interpretation. The move by EXAONE Finance to change the trunk rather than the branches reads as fatigue with that accumulation. It judges that replacing the body beats patching around it.

What the attention-free choice changes: the promise of linear scaling

The answer from EXAONE Finance is an attention-free architecture. It removes self-attention and aims for linear-time scaling. Length can grow while compute grows in proportion, and channels can grow while compute climbs on a manageable slope. When that promise holds, the daily weighing grows lighter. Teams can feed length without cutting and feed channels without thinning. Years of daily panels, and wide panels across many markets, can enter one model together. It is a choice that widens the menu of questions.

Unpack the word linear for a moment. Linear scaling means that doubling the length roughly doubles the compute. The place that used to quadruple under quadratic structure now doubles. The felt difference is large. An experiment that used to stretch from one day to one week now takes about two days. Shorter waits mean faster rotation of questions. Trials with longer context and wider channels multiply. Research speed is not the same as compute speed. Shorter waiting breeds more questions, and more questions breed deeper learning. The true worth of linear scaling is written on the question sheet rather than the compute sheet.

There are costs to name as well. An attention-free structure gives up the flexible long-range shortcut of self-attention. The direct expressway that joins two distant events in one hop disappears. Local mixing must stack layer upon layer to reach far places. Deep stacks do reach, but the middle path lengthens. A longer middle can thin the signal. So attention-free design is not only about removing attention but about filling the space left behind. What fills the temporal side and what fills the variate side is the whole design. The next two sections examine that filling one piece at a time. The craft of this work shows in how it fills.

Read this choice through the eyes of a desk. Production forecasting watches speed and memory and stability together. Hundreds of names must run before the open, and the next-day panel must be rebuilt after the close. Quadratic structures keep snagging that routine. Larger batches burst memory and longer windows miss deadlines. Linear structures loosen that daily breathing. The same machines see wider and run more often in the same hours. The research promise turns into a calendar story. This section is about timetables as much as compute tables.

Mixing time with causal 1D convolution

The device placed on the temporal side is a causal 1D convolution. It is a small window that slides along the time axis, as the name says. The word causal carries the weight. It reads only the past and the present, never peeking at the future. Peeking at the future is forbidden in forecasting. Scores look good in training when the future leaks, then collapse in production. Causal structure writes that taboo into the backbone. Windows that cover the right side and read only the left side stack in layers until flows from the distant past seep into the present. Many detours replace one expressway.

Unpack the feel of convolution for a moment. A small window reads local shape. Slopes over a few days, wobbles, and the way volume attaches to price all fall inside the window. Stacked layers widen the read. The first layer sees days, the next sees weeks, and the deeper ones see months. Longer breathing arrives with depth. There is a corner that fits the temporal structure of finance. Short jerks fall into short windows while regime slopes fall into deep layers. One device reads both breaths together. It reads steadily through fixed windows rather than reconnecting flexibly with attention weights. Steadiness turns into a virtue before noise.

Why is steadiness a virtue. The time axis of finance is noisy. A sharp rise on one day promises nothing about the next. Flexible connections overreact on such days. Weights rush to the spiking day, then wobble back as it cools. Fixed windows shake less in that swell. Since they read a fixed span in a fixed way, the noise of one day cannot shake the whole. Fixed windows can of course grow dull before sudden structural change. A shape never seen in the window takes more layers to read. It is a balance between sensitivity and dullness. Leading with causal convolution reads as more weight on steadiness before noise.

Length deserves one more thought. Convolution grows in proportion to length. It is the pillar that holds the linear promise on the temporal side. Years of daily curves can enter whole while compute climbs on a manageable slope. So the long breathing of regime turns can be watched without cutting. One-year and three-year and five-year windows can be fed side by side while asking where the living memory starts. Such experiments were postponed as dear under quadratic structures. On top of linear pillars they become daily work. The temporal mixing choice changes the menu of research questions at this point.

Mixing variates with group-aware pooling MLP

The device placed on the variate side is a group-aware pooling MLP. Pooling means pressing many values into one, and MLP means spreading the pressed values again through nonlinearity. The phrase group-aware is the seasoning. It means pressing in bundles before mixing everything. Names with similar texture, such as one sector or one market sleeve, gather first, then bundles mix with each other. It resembles hearing each choir part separately before hearing the full ensemble.

That order matters because financial entanglement has layers. Behind single-name moves sits a sector wind, and behind sectors sits the tide of the whole market. Mixing everything at once smears those layers. It grows hard to tell whether a signal is idiosyncratic to one name, or belongs to a sector, or belongs to the market. Gathering by groups first keeps the layers alive. Inside a group, idiosyncratic jerks are pressed together, and across groups the direction of the tide is read. It is mixing that preserves sources. Technique that blends without smearing.

Look deeper and group construction becomes the heart of the design. Groups drawn too large smear the layers, while groups split too fine lose meaning. Borrowing sector labels directly is easy, but labels never capture all of reality. Fresh themes and cross-border flows slip through old drawers. Learning groups anew on every run shakes training. A good compromise lives somewhere between fixed common sense and learned structure. Leading with the phrase group-aware reads as a direct engagement with that compromise. It declares an intent to keep bundle texture rather than lumping names together.

The MLP after pooling deserves equal thought. If pooling gathers, the MLP weaves. It kneads bundle signals through nonlinearity and reads pressure across names. Which bundle presses which, or how growth bundles react when the rates bundle moves, gets written into MLP layers. It reads steadily through learned weaving rather than reweaving weights on every step like attention. It shares the philosophy of the temporal side. Learned steadiness over flexible sensitivity. On loud panels that steadiness calms the wobble.

Through the eyes of a desk, this device also gives a thread for interpretation. Pressing by groups makes it easier to trace where a signal came from. Teams can ask which bundle moved behind a forecast, whether the force was single-name strength or sector wind or market tide. That source matters to portfolio builders. Idiosyncratic signals can carry weight, while market-wide tides call for hedges first. Mixing structure turns into explanation structure at this point. Well pressed pooling gives not only accuracy but a working language.

Growing strong on empty spans: masked context augmentation

Finance data is full of holes. Halts, listings and delistings, holidays, staggered disclosures by name, and monthly release calendars all punch holes. One name lacks yesterday, one indicator still lacks this month. Desks must live by filling those holes. Ignoring holes lets the model read emptiness as strange signal. Filling with zeros mistakes absence for zero, and forward filling peeks at structure that should stay hidden. Missingness is not an annoying exception but the daily life of finance.

Masked context augmentation brings that daily life into training. It deliberately covers part of the context during training and asks the model to read and predict through the cover. It covers not scattered dots but contiguous spans. It mimics streaky missingness that lasts days to weeks. Real financial holes arrive as streaks rather than dots. Holidays and halts and release lags come as spans. Training must mimic spans to resemble production holes. Closer mimicry means closer learning.

Why do contiguous spans matter. Scattered dots can be patched from neighbors. The mean of yesterday and tomorrow often suffices. Contiguous spans resist that trick. Distant edges thin the meaning of averages, and long spans turn patching into guessing. The model must read beyond edge pasting and use the surrounding context with signals from other channels. It learns to lift a missing name from its sector flow, and to fill a late indicator from other pieces of the market. It learns not how to patch holes but how to read through them. Density of augmentation sets intensity of training.

Seen over time, the direction of robustness appears. Models weak to missingness halt before holes. They hang on preprocessing that fills holes, and predictions wobble whenever preprocessing changes. Models strong to missingness pass through holes. They read what can be read through holes and admit what cannot be known. They shake less with the moods of preprocessing. From an operations view that gap is large. Lines must stand each morning even when disclosures lag overnight and some names halt. Masked context augmentation reads as a device that guards research scores and morning operations together.

Balance between too much and too little masking deserves one more pause. Covering too much thins the context itself. Panels full of covers teach little about what to read. Covering too little turns augmentation into decoration. Models halt again before long real holes. Length and frequency and placement of covers form the design. Some sleeves miss often and others miss rarely. Each sleeve carries its own texture of holes. Good augmentation resembles that texture. The closer the covers resemble the distribution of real missingness, the higher the production robustness. Designing augmentation means drawing a map of missingness.

Pretraining across six market sleeves

The pretraining slate reveals the ambition of this work. It feeds six sleeves together, equities and foreign exchange and commodities and crypto assets and fixed income and macro indicators. It learns the main dialects of finance rather than one or two markets. Each sleeve speaks differently. Equities speak in companies and sectors, foreign exchange speaks in nations and rates. Commodities speak in inventories and seasons, crypto assets speak in liquidity and sentiment. Fixed income speaks in maturity and credit, macro indicators speak in slow tides. Learning six dialects together trains an ear for the ensemble of markets.

Feeding many sleeves together is far from trivial. Clocks differ, amplitudes differ, and holidays differ. Crypto assets wail day and night while macro indicators whisper once a month. Aligning to equities leaves macro series full of holes, while aligning to macro smears equity texture. Normalization poses questions too. Which sleeve sets the eyes of the model. Fitting to loud sleeves buries quiet signals, while fitting to quiet sleeves bursts loud ones. Pretraining on six sleeves is construction work that holds such misalignment inside one model. Matching rulers and matching eyes is half the work.

The reason to feed them together is pressure across sleeves. When rates move, currencies shake, commodities react, and equity sectors rotate. One sleeve alone hides the source of pressure. Together the direction appears. The broader the pretraining, the better the model separates lonely signals of one market from ensemble signals of many. That separation shines when regimes turn. It can read where cracks opened first and where they spread. Broadly trained models wander less before unfamiliar regimes for this reason.

Think through the weight of the word foundation at this point. A foundation means a trunk learned once and used in many places. Teams place their own heads on a trunk hardened across six sleeves. One team adds domestic equities, another adds currency hedging. A sturdy trunk lets heads stay light. Local flavor can be reached with little data. A trunk tilted toward one sleeve makes heads elsewhere suffer. Reducing tilt is the heart of pretraining design. Feeding six sleeves evenly reads as a declaration to reduce tilt. It means learning the common language of finance rather than the dialect of one market.

Three tiers measured on FinVerse

Evaluation takes place on the FinVerse benchmark. It catches the eye because it has three tiers. It measures point-forecast accuracy with cross-sectional asset ranking and portfolio profitability together. Accuracy asks how closely curves were matched. Ranking asks whether risers were placed above fallers in proper order. Profitability asks whether a basket built on that order earned money. Higher tiers ask questions closer to practice. It is a staircase from laboratory rulers to desk rulers.

The reported result is simple. EXAONE Finance ranks first across all three tiers. It leads in accuracy, in ranking, and in profitability alike. Leading across three rather than one matters. Many models pair good accuracy with shaky rankings, and many pair good rankings with profits that never arrive. Taking all three together means curve reading and rank building and money making connect inside one trunk. Three crafts that used to play apart are strung together. So the phrase first across all three sounds larger than first alone.

Care is needed in reading numbers. First place means first inside that benchmark. Different periods and name universes and cost assumptions can reorder the board. The profitability tier is sensitive to assumptions in particular. Trading costs, turnover limits, and tradable hours reshape the look of money. So triple first places should not be read as a warranty for production. They should be read as a signal that three crafts connected inside the promises of the benchmark. Signals are clear while warranties are absent. Good reading separates signals from warranties.

The message of the sweep stays sharp all the same. A trunk designed apart answered together the questions that used to be asked apart. Linear structure without attention kept length and width, causal convolution and group pooling separated time and names, masked augmentation endured holes, and six-sleeve pretraining learned the ensemble, and that story folds into one line of triple first places. Component stories fold into a result story at this point. It is the pleasure of paper reading to watch blueprints fold onto score tables.

Reading from practice: missingness, length, and channels

Step outside the paper and picture a morning on a desk. Before the open, overnight indicators and holiday notices and halted names mix on quant screens. Some values miss, some arrive late, and some spike. Models must read that messy panel by a fixed hour. Late means late orders and wrong means losses. A trunk strong to holes, a trunk that sees long, and a trunk that sees wide all shine here. The three promises of EXAONE Finance meet three clocks of practice at this point.

Unpack the length problem on a desk. Reading regimes calls for length, but length bursts compute. So many teams cut. They watch only the latest year and drop what came before. Yet regime memory runs longer than one year. The tail of a hiking cycle and the start of a cutting cycle never fit inside one-year windows. Linear scaling pushes that cut backward. Three-year and five-year windows can be fed together while asking where living memory starts. It means testing memory length without watching compute nervously. Practice questions meet research promises at this point.

Channels tell the same story. Wide views reveal sector winds, but width bursts compute. So many teams thin names. They watch a few liquid leaders and drop the rest. Yet winds often blow first among dropped names. Strange jumps in neglected names, or joint stalls in one sector, arrive from outside the thin set. A widely seeing trunk brings that outside inward. It reads hundreds of channels at once and watches where pressure leaks. Attention-free structure with group pooling helps that eye. It means opening the eye of width without compute worry.

Missingness is the most daily problem. Some venue lags on one day and some indicator misses on another. Many teams start mornings on values filled overnight by preprocessing. When filling rules change, forecasts wobble, and wobbles shake trust. Models hardened on span missingness calm that shake. Since they read what can be read through holes, they shiver less with preprocessing moods. Holes too large must of course be admitted as unknown. Models that admit unknowns beat models that bluff on a desk. Robustness means working with holes rather than erasing them.

Remaining questions and next research

Good research leaves questions with answers. The first question left by EXAONE Finance concerns group design. If group-aware pooling worked, the next ask is how to build groups better. Fixed labels or learned bundles or blends of both remain open. Markets change and bundle texture changes with them. Yesterday's sectors never promise to equal tomorrow's. Keeping groups alive, and managing the shelf life of bundles, belongs to next research.

The second question concerns the breathing of time. If causal convolution beat noise through steadiness, the next ask is agility before sudden turns. How fast fixed windows read new shapes on days when regimes snap invites curiosity. Deep stacks see far pasts well but may see snapping presents late. Where to set the dial between steadiness and agility remains open. Different dials for calm days and wild days can be imagined too. Whether a trunk strong to noise stays fast to shocks, and how to hold both, forms the next chapter.

The third question lies outside evaluation. Triple first places on FinVerse give a clear signal, yet production money moves outside benchmarks. Trading costs and market impact and execution slippage gnaw profits. Rankings with high turnover often collapse before costs. Next come profitability with costs, and profitability with execution. It means measuring the strength of the bridge that turns ranks into money. It imagines a fourth tier of execution on top of accuracy and ranking and profitability. It extends one more practice step onto the laboratory staircase.

What can practitioners carry home right away. The most portable item is the attitude toward missingness. Before hanging only on preprocessing that fills holes, build the habit of hardening models by covering spans. Covering contiguous spans rather than scattered dots is the key feel. The next item is reviewing the habit of cutting length and width. Ask again how far length and channels can unfold under linear structures when they were cut because compute was dear. The last item is measuring with three rulers together. Place ranking and profitability rulers beside accuracy when accuracy alone used to suffice. Asking together whether curves were matched, whether ranks were built, and whether money followed sharpens the eyes of teams.

I close with the thought that lingers. Financial forecasting resembles enduring more than hitting. Regimes turn, holes open, and noise rushes in, yet the work reads what can be read, admits what cannot be known, and reads again the next day. EXAONE Finance showed a blueprint for an enduring trunk. It stands linear without attention, reads time through causality, presses names in bundles, rehearses holes in advance, and listens to six markets together. That the blueprint folded into one line of triple first places will linger long. While awaiting the next report on group shelf life and agility on snapping days and profitability with costs, I end this record here.