Levandor

crypto-shortterm — Short-Horizon BTC Prediction Research

12 min read

For Agents

Cited literature survey of short-horizon (seconds–5 min) BTC direction-prediction signals, assessed for a Polymarket 5-min BTC Up/Down system whose core edge is fair-value arbitrage — so signals are wanted primarily as an adverse-flow / toxic-fill filter and secondarily as candidate model features, NOT as a standalone predictor. Findings ordered most-actionable first. Every substantive claim carries a source. Full source artifact: /Users/levander/coding/pmv2/crypto_algo/research/short-term-crypto-prediction-strategies.md. Feeds crypto-shortterm-strategy-design, crypto-shortterm-data-sources, and crypto-shortterm-risk-management.

Survey of which very-short-horizon BTC signals are real enough to act on, which are popular-but-useless at ≤5 min, and the methodology traps that make high accuracy fail to translate into profit. Research date 2026-06-07; sourcing from primary scholarly APIs (arXiv, Crossref, GitHub).

The one constraint that changes everything: compute microstructure on Binance, NOT Polymarket

On Polymarket the tick rule and bulk-volume classification are near-random — 49.83% and 50.51% accuracy across 1.2B trades with 100% on-chain ground-truth aggressor direction (Qin & Yang 2026, “Polymarket-v1”). Positive trade-direction autocorrelation and concentrated market-making violate the mean-reversion assumption baked into those classifiers, so inferred VPIN diverges from true VPIN and OFI estimates are directionally biased on Polymarket. Any OFI / VPIN / CVD feature built from inferred aggressor side on Polymarket flow is unreliable. Rule: compute all aggressor-side microstructure signals (OFI, CVD, VPIN, lead-lag) on the clean Binance feed — which gives reliable aggressor flags — and use them to value and guard the Polymarket position. Only trust Polymarket-side flow if you can get ground-truth direction from the settlement layer, not heuristics. Source: Qin & Yang (2026), Polymarket-v1 Database — https://arxiv.org/abs/2606.04217

Actionable findings (most important first)

1. Strongest signal: multi-level Order-Flow Imbalance (OFI) on Binance

OFI — the signed aggregation of bid- vs ask-side limit-order arrivals, market orders, and cancellations — is the best-evidenced short-horizon driver of price and the #1 recommended signal for both the filter and the feature set.

  • Over short intervals price changes are mainly driven by OFI at the best bid/ask; the relation is linear, slope ∝ 1/depth, and the trade-volume relation is noisier than the OFI relation — Cont, Kukanov & Stoikov (2014), J. Financial Econometrics. https://doi.org/10.1093/jjfinec/nbt003 · preprint https://arxiv.org/abs/1011.6402
  • Out-of-sample fit improves with each additional book level → use multi-level OFI (MLOFI), top ~3–10 levels — Xu, Gould & Howison (2019). https://arxiv.org/abs/1907.06230
  • The genuinely forward (predictive, not just contemporaneous) OFI content lives at short horizons and decays in seconds — Cont, Cucuringu & Zhang (2023). https://arxiv.org/abs/2112.13213
  • Crypto, live-tested: microstructure features over short lookbacks explain 10–37% of 500 ms future returns across the most liquid BTC venues, with cost-inclusive PnL and a live maker experiment turning over $1.5M notional that beat a naive benchmark — Albers, Cucuringu, Howison & Shestopaloff (2021), “Fragmentation, Price Formation, and Cross-Impact in Bitcoin Markets” (the single most on-point public BTC result). https://arxiv.org/abs/2108.09750
  • Crypto, 1-second, Binance perps: OFI / spread / adverse-selection features show stable predictive importance across assets (BTC included); validated with taker and maker backtests — Bieganowski & Ślepaczuk (2026). https://arxiv.org/abs/2602.00776
  • Supporting: Sirignano & Cont (2018) universal price-formation map https://arxiv.org/abs/1803.06917 · Rahman & Upadhye (2024) OFI forecasting on Binance https://arxiv.org/abs/2411.08382

Clip the extremes — a naive linear OFI is most wrong when it looks strongest

Aggregate impact is non-linear, and extreme order-flow imbalance is NOT associated with large returns — at extreme imbalance the price is pinned; prices move when local flow is balanced (Patzelt & Bouchaud 2018, Phys. Rev. E, https://doi.org/10.1103/PhysRevE.97.012304 · preprint https://arxiv.org/abs/1706.04163). Keep OFI roughly linear and cap/clip the extremes. Also: don’t confuse contemporaneous explanatory R² with tradable foresight, and beware leakage from feature/label windows overlapping.

Implementation

OFI is O(1) per book update, computable in Rust from the L2 feed. As an adverse-flow filter it is close to ideal: strong OFI against the side you’re about to provide is the textbook signature of being run over.

2. The Market Maker’s Dilemma — heavy book imbalance toward your side is a WARNING, not a go signal

This is the operative finding for a passive / arbitrage bot, and it inverts the naive reading of order-book imbalance.

You get filled precisely when you're wrong

A live trading experiment on the Binance BTC perpetual found a negative correlation between maker fill likelihood and post-fill returns: the orders most likely to fill are the ones with the worst subsequent returns. Consequence — viable maker strategies often require a contrarian stance (counter-trading the prevailing order-book imbalance), and commonly-cited OBI-following strategies are highly unprofitable. Source: Albers, Cucuringu, Howison & Shestopaloff (2025), “The Market Maker’s Dilemma: Navigating the Fill Probability vs. Post-Fill Returns Trade-Off.” https://arxiv.org/abs/2502.18625

Operational reading: if we are about to be the resting bid and bid-queue imbalance is heavy and OFI is selling, our fill is likely toxic → widen, skip, or flip. This is core to the adverse-flow filter and to any Phase-2 market-making. Supporting queue-imbalance context:

  • Queue imbalance QI = (Vbid − Vask)/(Vbid + Vask) is a statistically significant one-tick-ahead predictor, much stronger for large-tick instruments — Gould & Bonart (2015). https://arxiv.org/abs/1512.03492
  • Volume imbalance materially improves directional prediction vs price alone — Lee (2024). https://arxiv.org/abs/2409.14157
  • Latency rank is the edge: OBI profit is allocated by being faster than the other OBI traders, not by absolute latency — Byrd, Palaparthi & Hybinette Balch (2020). https://arxiv.org/abs/2006.08682 → design these as fair-value/adverse-flow filters that don’t require winning the latency race.
  • Resting size near the touch is the easiest thing to spoof — treat real-book OBI as manipulable (Endo, Mizuta & Yagi 2025, https://arxiv.org/abs/2509.16912).

3. Binance as fair-value anchor + pick-off guard (don’t race the lead-lag)

Price discovery concentrates on the most liquid, lowest-fee venue (historically Binance for BTC); other venues lag by ms–seconds. We are explicitly NOT trying to win the cross-exchange latency race.

The feasible use

Treat the leading liquid spot venue (Binance) as the fair-value input for the Polymarket 5-min mark, and use a sudden Binance move vs a stale Polymarket quote as a pull-orders trigger (our resting order is about to be picked off). This sits squarely inside the arbitrage thesis and needs no sub-millisecond co-location.

  • Live cross-venue test: a leader-lagger network across liquid BTC venues, fee regime determines lead vs lag, tested live — Albers et al (2021). https://arxiv.org/abs/2108.09750
  • Price-discovery concentration (intraday/daily granularity): Baur & Dimpfl (2019), J. Futures Markets https://doi.org/10.1002/fut.22004 · Kapar & Olmo (2019), Economics Letters https://doi.org/10.1016/j.econlet.2018.10.031
  • The leader is not constant — time-varying, regime-dependent info flow between venues; don’t hard-code “Binance leads” (Vasiliauskaite, Lillo & Antulov-Fantulin 2021). https://arxiv.org/abs/2111.09057
  • Caveat: the raw lead-lag is heavily arbitraged, capacity/latency-bound, and depends on what reference price Polymarket actually settles against.

4. Volatility / cascade circuit-breaker (gate trading; predicts magnitude, not direction)

Volatility clustering is the best-established stylized fact in finance and strongly present in BTC — but it tells you how big, never which way. Using vol as a direction signal is a category error. Its value is as a conditioner, sizer, and skip filter.

Circuit-breaker conditions

Stop quoting / widen drastically on: RV spikes, liquidation-print bursts, spread blowouts, and scheduled volatility — full-hour algorithmic activity bursts and US macro releases (NFP / CPI / FOMC), where fills are maximally toxic. Volatility timing is partly scheduled (Wątorek et al 2023, Chaos, https://doi.org/10.1063/5.0165635). Also size by inverse vol, and feed RV as a feature that modulates every other signal.

5. Secondary confirmation: VPIN-style flow toxicity + CVD slope on Binance

The trade leg of order flow (signed executed volume; CVD = running sum) is real but noisier and less robust than book-event OFI — already established in Cont-Kukanov-Stoikov (2014). Use as secondary confirmation, computed on Binance.

  • VPIN (volume-synchronized probability of informed trading) is the canonical adverse-flow construct — high VPIN flags toxic order flow that precedes adverse selection for liquidity providers — Easley, López de Prado & O’Hara (2012), Review of Financial Studies. https://doi.org/10.1093/rfs/hhs053 (treat VPIN as a toxicity/illiquidity gauge, not a crash oracle — its crash-prediction claim is contested.)
  • BTC-specific: OFI + preponderance of aggressive traders + widening spread predict jumps, and jumps cluster — Scaillet, Treccani & Trevisan (2017). https://arxiv.org/abs/1704.08175
  • CVD is path-dependent and drifts; needs careful windowing/normalization or it just tracks cumulative volume.
  • Reminder: do not compute toxicity from inferred direction on Polymarket flow (§ top warning) — Binance only.

Sub-minute momentum vs mean-reversion (conditioned feature only)

Two competing very-short autocorrelation effects (continuation from flow persistence vs reversion from liquidity replenishment); which dominates is horizon- and regime-specific and sign-flips (Hu & Zhang 2025, https://arxiv.org/abs/2505.17388). Per-trade edge is tiny vs costs. Useful only as a conditioned feature (short-horizon return interacted with vol and OFI) and as part of the reversal-aware fill filter — never as a naive “follow the last-minute trend” rule. Persistence evidence: Sirignano-Cont (2018). Reversion-at-the-touch is the Market-Maker’s-Dilemma “Reversals” result (Albers et al 2025).

Wrong timescale — do not use as 5-min direction features

  • Funding rates / perp basis — positioning/carry variables on an hours-to-days horizon; essentially constant per 5-min bar, and contrarian/regime-dependent even at native horizon. At most a slow squeeze-risk context flag. (Cao et al 2026 https://doi.org/10.2139/ssrn.6795783; Matsui et al 2022 https://doi.org/10.1109/ICBC54727.2022.9805528.)
  • On-chain / whale flows — empirically a 1–6 h horizon (USDT exchange inflows predict returns; whale tx → next-day vol). BTC block time ~10 min + attribution lag make it useless at ≤5 min. (Chi, Chu & Hao 2024/25 https://arxiv.org/abs/2411.06327; Herremans & Low 2022 https://arxiv.org/abs/2211.08281.)
  • Liquidation-cascade direction — the phenomenon is real and material (BitMEX: 3.51% long / 1.89% short of outstanding futures liquidated daily, ~60× leverage — Cheng et al 2021, https://doi.org/10.1080/00036846.2021.1922597), but as a directional predictor it is folklore with no rigorous public backtest. It is reactive not predictive (the move has largely happened by the time prints arrive), rare (tiny-sample overfitting magnet), and direction-ambiguous (continuation vs exhaustion snap-back). Use only as a risk flag in the circuit-breaker, never to predict per-bar direction. Early-warning detectors get positive lead-time only in high-SNR regimes (Hiremath & Hiremath 2026, https://arxiv.org/abs/2604.20949).
  • Naive last-minute momentum — sign-flips by regime (Hu & Zhang 2025); not a reliable standalone direction call.

Methodology traps (read before trusting any backtest)

High accuracy ≠ profit after spread/fees

State-of-the-art DL has high forecasting power that does not necessarily correspond to actionable trading signals; accuracy/F1 fail to assess LOB forecast quality — evaluate the probability of correctly forecasting a complete transaction instead (Briola, Bartolucci & Aste 2024, https://arxiv.org/abs/2403.09267). Across FI-2010, NASDAQ and a Bitcoin dataset, predictability has declined over time (−6.68 F1) and, when trends are defined relative to the average spread (the real cost), performance deteriorates sharply (Berti & Kasneci 2025, “TLOB”, https://arxiv.org/abs/2502.15757). A 70–78% mid-move accuracy number alone is not evidence of profit. Demand a cost-and-spread-aware backtest (maker/taker fees, queue position, slippage).

Look-ahead / label leakage

An inadequately defined target price process renders predictions meaningless by incorporating past information, and the common 3-class label conflates volatility and direction (Lee 2024, https://arxiv.org/abs/2409.14157; horizon-bias-removing labels in Berti & Kasneci 2025). Use strictly future, non-overlapping labels; never let the smoothing/normalization window touch the label.

Backtest overfitting / multiple testing — the default, not the exception

Try enough OFI/QI windows and you will find a great in-sample Sharpe by chance. Deflate the Sharpe for selection bias, non-normality, and short samples (Bailey & López de Prado 2014, https://doi.org/10.3905/jpm.2014.40.5.094), and demand t ≈ 3+ for a new signal, not the lax t ≈ 2 (Harvey, Liu & Zhu 2016, https://doi.org/10.1093/rfs/hhv059). Pre-register a small feature set; use walk-forward / purged CV.

Domain mismatch + latency-rank reality

Polymarket ≠ a crypto exchange — compute microstructure on the clean Binance feed (§ top warning). And OFI/QI/lead-lag edges are won by relative speed (Byrd et al 2020) — design them as fair-value and adverse-flow filters that don’t require winning the latency race.

Adverse-flow / toxic-fill filter (primary):

  1. Multi-level OFI on Binance, clipped at extremes — the core toxicity signal.
  2. VPIN-style toxicity bucket + CVD slope on Binance — secondary confirmation.
  3. Order-book imbalance read with the contrarian sign — heavy OBI in favor of the side we’re about to provide ⇒ likely toxic fill ⇒ widen/skip/flip.
  4. Binance jump vs stale Polymarket quote ⇒ pull resting orders (pick-off guard).
  5. Volatility / cascade circuit-breaker — RV spike, liquidation-print spike, spread blowout, or scheduled macro / full-hour bursts ⇒ stop quoting / widen.

Candidate model features (secondary, cost-aware validation only): OFI (multi-level), QI, CVD/VPIN, recent RV (as conditioner), short-horizon return interacted with vol, a Binance-lead basis term. Volatility as a pervasive conditioner/sizer, never as direction.

Validation discipline (non-negotiable): cost+spread-aware backtest with realistic queue/slippage; future-only non-overlapping labels; walk-forward / purged CV; deflated Sharpe; t ≈ 3 hurdle; signals computed on the clean exchange feed, not on heuristically-classified Polymarket flow.

Source artifact

Full cited report (with evidence grades per claim): /Users/levander/coding/pmv2/crypto_algo/research/short-term-crypto-prediction-strategies.md

Graph View

Backlinks