Image-scaled price windows, the representation behind chart-reading networks, tested beside the rankers HydraTrade already runs.
In 2023 Jiang, Kelly and Xiu showed that a convolutional network reading price charts ranks US stocks better than any classic trend signal they compared it with.1 Their own ablation points somewhere quieter than the pictures. A one-dimensional network fed the same numbers, scaled the way a chart scales them, did just as well, and the same network fed volatility-scaled returns did nothing. The useful part was the representation.
This report asks whether that representation adds anything to the rankers HydraTrade already runs, in the crypto book that trades in paper and in twenty years of the 500 most traded US stocks. Seven arms were registered before any was run. The image columns help a little in stocks at a five-day hold, the horizon where the paper is strongest, and nowhere else. A network that sees only the chart ranks worse than the trees in both markets. The trees had already read most of the picture.
Technical analysis has always had a mixed reputation in finance, and the evidence on it is genuinely mixed.
Lo, Mamaysky and Wang smoothed prices with kernel regression, detected head-and-shoulders and similar patterns by algorithm, and found that their appearance shifts the distribution of the next returns a little.2 Brock, Lakonishok and LeBaron found that simple moving average and breakout rules earned money over nearly a century of the Dow.3 Sullivan, Timmermann and White then counted the thousands of rules a researcher could have tried. The best of them survived that correction in the original sample and earned nothing in the decade that followed.4 The pattern of that literature is a real but small effect, easy to overstate.
Jiang, Kelly and Xiu changed the question.1 They drew every US stock's last 5, 20 or 60 days as a black and white bar chart, trained a CNN on 1993 to 2000 to predict whether the stock would rise, and held it fixed through 2001 to 2019. Sorted into deciles, its weekly long-short portfolios earned annualised gross Sharpe ratios as high as 7.2 with equal weights and 1.7 with value weights. Their Table IX holds the result this report builds on. A 1D CNN reading the same five days as numbers, each price written as its position between the window's lowest low and highest high, scored 7.20 against the 2D network's 7.15. Fed returns divided by their volatility, the same 1D network scored −0.13.
Two other numbers from the same paper set the expectation for this one. Restricted to the 500 largest stocks, the five-day strategy's equal-weight Sharpe falls from 7.2 to 1.02, and it turns over about 700 % of the portfolio a month (their Table IV). Murray, Xia and Xiao, using a different network on past returns, also find the effect strong among the largest 500 stocks.5 HydraTrade's stock universe is exactly that kind of liquid set, and its crypto book is smaller still.
The features HydraTrade's trees read are returns over lags from one day to six months, volatility, dollar volume, distance from a moving average, the recent high-low range and, in crypto, positioning data from the derivatives market. The closest cousin of image scaling is the position of the close inside a longer range, the quantity behind George and Hwang's 52-week-high effect.6 Nothing describes the shape of the last week. So the question is fair. If what the network extracts is mostly short-term reversal,7,8 sharpened by knowing where each day closed inside its range, a tree that already sees reversal and range position should gain little. If it is genuinely new, as the paper argues, the columns should add.
Every arm, parameter and pass rule below was fixed in a register before the first arm ran, and every later change is logged there with its reason.
| Channel | Value for each day of the window |
|---|---|
| open, high, low, close | (price − L) / (H − L) |
| moving average | the W-day mean close on the same scale, free to leave [0, 1] as the average line leaves a chart |
| volume | dollar volume over the window's largest |
Windows are five and twenty days, the paper's I5 and I20, which gives 150 values per asset and day. For the trees each value is ranked across the universe on the decision date, the convention every other feature follows. The CNN reads the unranked 20 × 6 window. In stocks a decision on day D uses days up to D. In crypto the decision at midnight UTC uses complete days up to D − 1.
The CNN is deliberately plain. Two 1D convolutions (6 to 32 to 64 channels, kernel 5) are followed by mean and max pooling over time, dropout and a linear head, about 11,400 weights in all. It reads only the window, with no feature row and no view of other assets, as in the paper. It learns the trees' target, the within-date rank of the forward return, with the optimiser and early stopping of HydraTrade's earlier sequence models. Three seeds are trained and their within-date ranks averaged, since one seed of a small network is a draw and not a model.9
The pass rules came from report III. A tree arm is rescaled to the baseline's volatility, and its weekly gain needs a Newey–West t above 2 and a positive mean in both halves of the sample. In crypto it must also grow faster at the largest leverage the risk budget allows. A CNN arm is regressed on the tree sleeve it replaces, and the intercept needs a t above 2, which asks whether the network earns anything the tree does not. The stated prior was modest. Report I had found one-day reversal losing money in the crypto book, and the paper's strongest pattern looks like a reversal.
Both gates passed to the last decimal, so the arms below differ from their baselines only in what the register says they change.
With the image columns, the crypto book's Sharpe ratio moved from 1.74 to 1.85, a gain of 2.4 basis points a week at equal volatility with a t of 1.15. Nearly all of it came in the second half of the 233 weeks (0.3 and 4.6 bp). Sleeve A with and without the columns correlates at 0.95, so the trees ranked coins almost as before. The leverage test passed (22.5 % a year against 20.2 % at half the backtest edge) but the t did not, and the register asked for both.
This is the arm that passed. The Sharpe ratio rose from 0.36 to 0.51, the gain was 2.1 bp a week with a t of 3.02, and both halves were positive. It is also the one place where the paper's evidence predicted something, a one-week horizon in US stocks.
It deserves two cautions. The deflated Sharpe ratio,10 counting the 15 stock configurations of report II and this report's three, puts the probability that the book's true Sharpe is above zero at 0.03. The columns improved a weak book without making it a strong one. And the gain is front-loaded. Figure 4 climbs steeply until about 2015, moves sideways until about 2020 and then recovers some ground. Split at the midpoint, that is 3.3 bp a week in the first half and 0.9 in the second. Anomalies in liquid US stocks have weakened in the high-volume era,11 and published ones decay after publication,12 though twenty years of one arm cannot say which, if either, is at work here.
At a monthly hold the columns did nothing (t −0.21, halves of −1.9 and +1.5 bp). The paper finds its monthly image strategy earns most of its return in the first five days after the image (Table III there), so a fading signal over a longer hold is what one would expect.
Taking the trees away and letting the network rank on the window alone did worse in both markets.
The CNN sleeve earned a Sharpe ratio of 0.43 against the tree sleeve's 0.95. Its returns correlate with the tree's at 0.41, and after regressing them on the tree sleeve the intercept is 1.5 bp a week with a t of 0.08. It ranks coins differently and earns nothing for it.
One side effect needs reporting. Swapping a strong sleeve for a weaker, less correlated one made the whole book calmer. Its worst drawdown fell from 13.3 % to 7.7 %, so the risk budget would allow more leverage, and at that leverage, with the backtest's edge halved, the book would grow 26.2 % a year against 20.2 %. At a fixed 2× on the full backtest edge it grows less, 35.5 % against 43.9 %. This is diversification paid for with return, and it is why the register judged CNN arms on the residual rather than on the book. A cheaper diversifier might do the same job, which would need its own registered test.
The stock CNN lost money, a Sharpe ratio of −0.71, with a residual of −6.8 bp a week and a t of −3.97. A result that confidently negative can mean a flipped sign somewhere, so after the run every saved model was rescored on its own test year, with nothing refitted. This check was not in the register and is reported as such.
There was no flipped sign. The network ranks in the right direction, only weakly. Before costs its sleeve earns 3.6 bp a week with a Sharpe ratio of 0.42 (t 2.00), against 14.5 bp and 1.07 for the tree. The stock simulation, inherited from report II, charges every daily tranche a full round trip whatever the model holds, 9.7 bp a week here. That fixed charge takes the tree from 14.5 to 4.8 and the CNN from 3.6 to −6.1, and it is also why the residual test, run on net returns, lands so far below zero.
The paper's 2D network, in the 500 largest stocks from 2001 to 2019, earned a gross equal-weight Sharpe of 1.02. This one earned 0.42 from 2006 to 2026. The two differ in more than architecture. The paper trains once on the whole market, small stocks included, where the pattern is strongest, and this network only ever saw the 500 most traded names. The gap cannot be pinned on any one of these.
A good representation helps most when the model does not already have something close to it.
The trees were never blind to price shape. They see reversal at one, two and five days, the recent range and the distance from an average, and boosted trees are very good at carving such inputs into the interactions that matter. On tabular data they routinely match or beat deep networks,13 and in Gu, Kelly and Xiu's horse race of machine learning methods for US returns, networks finished only slightly ahead of trees, and both drew most of their gains from a small set of price-trend, liquidity and volatility signals.14 Image scaling hands a tree a sharper view of where each day sat inside its week. That is new, but only a little, and the passing arm measures how little, about 2 bp a week.
Liquidity explains the rest of the shape. Short-horizon reversal is largely pay for supplying liquidity,15 and it is concentrated in illiquid stocks.16 A universe of the 500 most traded names is where that pay is thinnest, and the paper's own numbers fall sevenfold when it moves there. Crypto behaves differently again. Its cross-section is described well by market, size and momentum factors,17 and in this book one-day reversal had already lost money, so a representation built for the shape of the last week had little to grab.
The two CNN arms point the same way from the other side. A network that sees only the window gives up everything the trees know about longer trends, volatility and, in crypto, positioning. What it finds alone is real, positive in 17 of 20 years in stocks, but it is a fraction of what the trees find. The picture is worth having as a few extra columns and not as a replacement for the rest of the evidence.
Nothing in the live book. The crypto arms failed, so the paper-trading book keeps its current features.
The stock result passes its registered test and goes on record. The stock book itself stays a research object, with a deflated Sharpe ratio of 0.03, so nothing is deployed from it.
The calmer IX2 book is noted as a question, whether a weak and loosely correlated sleeve is worth carrying for its diversification, to be registered on its own if it is pursued.
It does not show that chart images are useless. Only one small 1D network was tried, untuned by design, and no 2D network at all. The paper finds its 2D network no better than its 1D one, so the missing tuning matters more than the missing pictures. A larger network, or one trained on the whole market and then applied to the liquid set as the paper did, could do better.
It does not show the effect is absent in small stocks. They were left out on purpose, because this programme cares about what it could trade, and the literature puts most of the effect there.
The cost model is a fixed convention. If the CNN trades less than it charges, the net figures are too harsh on it, and if more, too kind. The before-cost figures in Figure 6 do not depend on that choice, and they already rank the tree well above the CNN.
Multiple testing was counted, but only as well as counting allows. Five arms were tested and one passed at t 3.02. Harvey, Liu and Zhu argue a new factor should clear a t near 3 given how many have been tried,18 and this one only just does.
The crypto sample is 233 weeks on one venue. A small edge in either direction there would take years more data to see.
Data. Crypto daily bars are UTC days built from the venue's hourly bars, with dollar volume summed from hourly volume times close. Stock bars are split-adjusted daily prices of US common stocks from 1998, with dollar volume from the unadjusted close times shares traded. The stock universe on each date is the 500 securities with the largest trailing dollar volume, chosen point in time.
Walk-forward. The crypto book uses report III's nine 180-day windows over 233 weeks. The stock book uses twenty 365-day test windows from October 2006 to September 2026, 1,043 weeks. Each stock model trains on everything before its window, holds out the last 15 % of training dates for early stopping, and leaves an embargo of twice the holding period between training labels and test dates.19 The crypto models follow report III's equivalent rules.
Portfolios. Stocks are held in equal-weight decile legs, long the top and short the bottom, as overlapping daily tranches over the holding period, at 10 bp per leg per round trip. Crypto runs report III's v2 book of four sleeves with costs charged on net position changes. Everything is measured at 1× gross.
Statistics. Sharpe ratios are annualised from weekly returns. t statistics use Newey–West variances with Bartlett weights and 4(n/100)2/9 lags.20 Gain tests rescale the arm to the baseline's weekly volatility before differencing. Residual tests regress the CNN sleeve's weekly returns on the tree sleeve's. The deflated Sharpe ratio assumes a dispersion of 0.5 in annual Sharpe across trials. The leverage budget is the largest gross leverage whose bootstrapped chance of a 50 % drawdown within three years stays under one in ten, with the backtest's mean return halved.
Reproducibility. Every fitted model is saved for every arm and window, and each arm records the code version and a hash of every file that defined it. Both gates reproduced earlier series to within about 10−16 a week. The stock CNN run was interrupted once near its end and resumed from its saved models, which is equivalent because each window's training starts from its own seeds.
Three changes were made after registration, each logged with its reason and none touching a model, its data or its settings. A missing method stopped the first IX2 run before it fitted anything. The image-column builder was made faster, with identical output. The stock CNN's training loop was rewritten for speed and checked against the original to give the same predictions on test data.
Research publications of Ciya Labs. Not investment advice; backtested and simulated results do not predict future returns. Imprint · Privacy