Skip to content

Universal Price Formation

Definition

Universal price formation is the hypothesis that the mapping from the state of the limit order book (prices, sizes, event history) to the next price move is not asset-specific: a model trained on pooled data from many stocks performs at least as well — and often better — than models trained stock-by-stock, including on stocks it has never seen during training.

Formally, if \(F_i(X_t; \theta_i)\) denotes the price-formation map for asset \(i\) with parameters \(\theta_i\), the universality claim is that there exists \(F(X_t; \theta)\) (single \(\theta\), no \(i\)) which approximates \(F_i\) for every asset \(i\) in a broad class.

Historical context

The universality idea has two lineages:

  1. Universality in scaling laws. Since the late 1990s, empirical work (Kyle-Obizhaeva invariance, Patzelt-Bouchaud nonlinear impact scaling, Bouchaud-Farmer-Lillo propagator model) documented that many microstructure statistics — returns distributions, price-impact curves, volume-duration relations — exhibit cross-asset universality after appropriate rescaling.
  2. Universality in the input-output map. Sirignano & Cont (2019) go further: not just the statistics but the full function mapping LOB history to next-price-move is stock-agnostic. This stronger claim is what "universal price formation" usually means now.

Evidence

Deep-learning evidence (Sirignano-Cont 2019)

Training a single 3-layer LSTM on pooled Nasdaq-L3 data from ~500 US stocks (Jan 2014–May 2015), then testing on:

Test set Result
Same 500 stocks, later 3 months Beats stock-specific LSTMs on most stocks
500 new stocks never seen in training Matches/beats stock-specific LSTMs (avg +0.2%)
Same stocks, 18 months after training No significant accuracy decay
Longer training windows (1 mo → 19 mo) Accuracy grows monotonically

No sector- or tick-size-specific pre-partitioning was beneficial.

Linear-model evidence (Cont-Cucuringu-Zhang 2023)

A sparser, linear follow-up: cross-impact-ofi-equity-markets shows the integrated OFI relation holds with similar \(R^2\) structure across the Nasdaq-100 universe. The coefficient is not constant across stocks, but the form of the relationship is universal.

Earlier linear evidence (Cont-Kukanov-Stoikov 2014)

price-impact-order-book-events documents that the linear OFI → price-change slope \(\beta\) follows \(\beta \propto 1/\text{depth}\) with no free parameters across 50 NYSE stocks — a universality up to a depth rescaling. Sirignano-Cont's result absorbs even that rescaling into the learned model.

How it works (mechanism)

The implicit claim is that market participants — HFTs, market makers, institutional algorithms, retail order routers — deploy roughly the same classes of strategies across stocks. Their order flow therefore exhibits the same statistical structure (queueing behaviour, cancellation patterns, depth decay) regardless of which ticker they operate on. The LOB → next-price-move map is a function of this collective behaviour, so if the behaviour is universal the map is too.

Consequences of this view:

  • Pooled data is strictly better than sharded data for training. A universal model sees 500 stock-years where a stock-specific model sees 1.
  • Newly-listed stocks inherit the behaviour from the universal training set on day one — no warm-up period needed.
  • Regime changes that affect behaviour (e.g. Reg NMS, decimalisation, HFT proliferation) are the real non-stationarities to worry about, not calendar time per se.

Open questions

  • Cross-market universality. Does the same map work for futures, FX, crypto, single-name options? Tentative evidence suggests no (tick-size regimes differ, latency differs, participant mix differs), but this has not been rigorously tested.
  • Regime boundaries. Sirignano-Cont's 18-month stability spans a calm period (2015–2017). How stable is the map across 2020, 2022, or a major market-structure reform?
  • Mechanism isolation. What subset of the input features actually drives the universality? Can we recover a low-dimensional "universal core" from a trained LSTM?
  • Universality for levels beyond the top. Evidence is cleanest at the top of the book. Deep-book universality (levels 5–10) is argued but less rigorously tested.

Connections