How Learnable Is BTC Fair-Market Value from HLCOV?

The premise is deliberately narrow. Rather than asking whether Bitcoin can be traded profitably, the study asks whether the next minute's fair-market value carries any predictable signal in the most basic data a market emits: the high, low, close, open, and volume of each one-minute bar. Fair-market value is pinned to a concrete target — the next bar's typical price, (H+L+C)/3 — with an option to swap in the next-bar close instead.

The point of framing it as learnability is to strip away the machinery of a trading system and isolate one measurable thing. If a graded set of models, from a linear baseline up to a sequence network, cannot extract a signal from HLCOV, then over this feature set learnability is effectively zero. The experiment is built to give that answer cleanly rather than to flatter any particular method.

The data is BTCUSDT one-minute klines pulled from Binance's public dump at data.binance.vision, which serves the same shape of one-minute history as the originally requested Kaggle dataset but without any authentication or rate limits. The default fetch is the six most recent complete months, with each monthly tarball around ten megabytes compressed.

Features are the current bar's HLCOV, the last N lagged HLCOV bars (default lookback of 20), and a handful of standard derived signals: returns over the lookback window, range, a volume ratio, and return-volatility. The models form a ladder chosen to answer distinct questions — naive baselines to set the bar, ridge and lasso to test whether a linear signal exists, random forest and HistGradientBoosting to catch tree-friendly non-linearities, an MLP to confirm the gradient-boosted models are not underfitting, and a PyTorch sequence-to-one LSTM to test whether the raw ordering of the five-channel time series carries information the engineered features miss.

Evaluation uses a strict chronological train/test split: the final 20% of the window is held out with no shuffling, and features at bar t only ever use bars up to t while the target is bar t+1, so there is no lookahead leak.

The most important measurement decision in the study is what not to trust. On the price level, R² is inflated by autocorrelation to the point of meaninglessness — a model that simply predicts the next price equals the current price scores around 0.999. A near-perfect R² here says nothing about whether anything was learned; it merely restates that adjacent minute prices are close together.

For that reason the study elevates directional accuracy — the sign of the next-bar return — as the metric that actually measures learnability, with MAE and RMSE alongside it. Getting the direction right is the part that requires real signal; getting the level approximately right is nearly free.

The answer from version 1 is a flat one. Across every model in the ladder — naive, ridge, lasso, random forest, HistGradientBoosting, MLP, and LSTM — next-bar directional accuracy came in at roughly 0.5. No family beat the coin flip, and nothing beat the naive baseline on the metric that matters.

That uniformity is itself informative. When linear models, tree ensembles, a dense net, and a sequence model all land at chance, the shortfall is unlikely to be an underfitting artifact of any one method. Over this feature set and horizon, the signal simply is not there to be found.

Rather than declaring the question closed, version 2 responds to the flat v1 result by attacking both halves of the setup. The feature library expands roughly tenfold to about 89 quantitative features: realized-volatility estimators (Parkinson, Garman-Klass, Rogers-Satchell, Yang-Zhang), oscillators (RSI, MACD, Williams %R, Stochastic, CCI), trend and band indicators (ADX, Aroon, ROC, Bollinger, Donchian, Keltner), volume and microstructure signals (OBV slope, Chaikin money flow, MFI, Amihud illiquidity, Roll spread, VWAP deviation, taker-buy ratio, Kyle's lambda, VPIN), statistical descriptors (Hurst exponent, variance ratio, autocorrelations at lags 1 and 5, skewness, kurtosis, approximate entropy), and quant ratios (rolling Sharpe, Sortino, Kelly, max drawdown, Calmar, Ulcer), aggregated across 5-, 15-, and 60-minute scales.

The definition of fair value is also revisited rather than assumed. Version 2 sweeps seven FMV targets — close, typical, VWAP, the (H+L)/2 midpoint, the open-close midpoint, the geometric mean, and a five-bar-ahead mean as a longer-horizon target. Two deep models join the LSTM: a 1D CNN and a small Transformer encoder (two layers, four heads, d=64) over the lookback feature sequence. Permutation feature importance on the HistGradientBoosting model is used to see which of the many features actually move the held-out error. Outputs land in a separate results_v2 directory.

The project is scrupulous about its own boundaries. No fees, slippage, or spread are modelled, so a directional accuracy near 0.51 on an average move of a few basis points would not be tradeable even if it were real. The evaluation runs over a single contiguous window rather than walking forward across distinct regimes such as choppy 2022 versus the 2024 bull, which is what a genuine backtest would require. This is explicitly a learnability check, not a trading strategy.

The engineering matches that modest scope: the whole run is designed to finish in roughly five minutes on a laptop, training on up to 300k rows, niced down so it does not pin the CPU. The value on offer is not a model to deploy but a clean, reproducible, leak-free answer to a specific question — and an example of measuring the answer with the metric that cannot be gamed by autocorrelation.