shalev396/eu-stock-forecasting

EU Stock Market Forecasting

Forecasts the four European stock indices of R's classic EuStockMarkets dataset (DAX, SMI, CAC, FTSE, 1,860 business days, 1991-07-01 to 1998-08-14 on a synthetic calendar) from any cutoff date, with a classical time-series model that only sees the history up to that date. Five forecasters were compared on every index (ARIMA, Holt-Winters, VAR, a Keras MLP and a Keras LSTM) against the naive "tomorrow = today" baseline. Per index, the deployed model is the classical model with the fewest fitted parameters among those whose validation error is within 2 % of the best: Holt-Winters (4 parameters) on DAX, SMI and FTSE, and VAR(10) (164 parameters) on CAC, where nothing smaller comes close. On the DAX test window (last 100 business days), the deployed Holt-Winters has a one-step-ahead MAPE of 1.047 % (RMSE 72.9 index points). The honest headline: on DAX all three classical models land within 0.02 MAPE points of naive persistence (1.048 %). Daily index closes are close to a random walk, and nothing here beats that by a meaningful margin. Numbers are from the run of 2026-09-25 (metrics.json).

Model

• Input: cutoff (a YYYY-MM-DD date), horizon (1-120 business days) and index (DAX, SMI, CAC or FTSE).
• What it ships: config.json (the deployed method per index, plus every classical method's settings: ARIMA order + coefficients, VAR lag, Holt-Winters trend) and eustocks.csv (the ~82 KB dataset the forecasts run on). There are no pickled fits. Each request refits on the history up to the cutoff in well under a second:
  • Holt-Winters (deployed for DAX, SMI and FTSE), additive trend and no seasonality, with 4 parameters (smoothing level and trend, initial level and trend). It is refit per request, and predict(..., method="holt_winters") selects it. It is also the model behind the Space's bring-your-own-CSV tab (predict_series).
  • VAR(10) (deployed for CAC): a vector autoregression on the four index levels together, 10 lags picked by AIC (max 12), refit by OLS on every request. It has 164 coefficients (10 lag matrices of 4x4 + 4 intercepts). Select it with method="var".
  • ARIMA(p,1,q), with the order picked by AIC on the fit window per index (DAX (2,1,3), SMI (2,1,3), CAC (3,1,3), FTSE (3,1,2)). Serving keeps the trained coefficients and re-runs the Kalman filter over the history up to the cutoff, because MLE refits of these near-cancelling ARMA orders take 1-4 s and often do not converge. Select it with method="arima".
• Output: a JSON dict {model, history, forecast, actual, naive, mae, naiveMae}. history holds the last 250 closes up to the cutoff. forecast, actual and naive cover the horizon business days after the cutoff: the model's forecast, the real closes (empty past 1998-08-14), and the last close carried forward. mae and naiveMae are the mean absolute errors against actual (null without actuals).
• Dates are synthetic. The R ts object has no calendar (start 1991.496, 260 observations a year), so row 1 is dated 1991-07-01 and each following row is the next weekday (Mon-Fri, no holidays). Forecast dates continue that calendar. The order and spacing of the data are real; the exact dates are approximate.

Usage

from huggingface_hub import snapshot_download
import sys
path = snapshot_download("shalev396/eu-stock-forecasting")
sys.path.insert(0, path)
import model
predictor = model.load(path, device="cpu")
out = predictor.predict("1997-10-01", 30, "DAX")                  # deployed Holt-Winters
print(out["model"], out["forecast"][:3], out["mae"], out["naiveMae"])
predictor.predict("1997-10-01", 30, "DAX", method="var")           # or "arima" / "holt_winters"

• Space API (free): POST https://shalev396-eu-stock-forecasting.hf.space/gradio_api/call/predict with {"data": ["1997-10-01", 30, "DAX"]}. See the Space.
• Inference Endpoint: handler.py makes this repo deployable (Deploy -> Inference Endpoints, CPU). Request body: {"inputs": {"cutoff": "1997-10-01", "horizon": 30, "index": "DAX"}}.

Training

Chronological split per index, never shuffled: fit = 1,660 days (1991-07-01 .. 1997-11-07), validation = 100 days (.. 1998-03-27), test = the last 100 days (1998-03-30 .. 1998-08-14). The ADF test finds a unit root in every index level (p >= 0.76) and none in the first differences (p < 0.001), which is why ARIMA uses d = 1. All parameters are estimated on the fit window only: the ARIMA(p,1,q) grid (p, q <= 3, 15 fits per index, ranked by AIC), Holt-Winters, the VAR lag (by AIC), and the neural nets. The nets are an MLP 64-32 on 20 lagged closes (3,457 parameters) and an LSTM(64) on the same 20-day window (16,961 parameters), trained with Adam + MSE on MinMax-scaled closes (scaler fit on the fit window) and early-stopped on validation.

Evaluation is one-step-ahead walk-forward with frozen parameters: each forecast sees the actual closes up to the previous day. This is the setting of R's forecast::accuracy(), whose table (ME, RMSE, MAE, MPE, MAPE, MASE) is reproduced below. The deployed method per index follows a rule fixed before the test window was looked at (see Experiments): fewest fitted parameters within 2 % of the best validation MAPE. The test window is only reported. A multi-step backtest then replays the Space itself: from 9 origins, one every 10 days of the test window, model.forecast sees the history up to the origin and predicts 20 business days ahead.

Hardware: CPU (a shared 20-thread desktop). The full notebook run took about 3.5 min of wall time, 168 s of it for fitting (ARIMA grids and Keras). Full code: training/ · Colab

Evaluation

Deployed model (Holt-Winters, additive trend) on the DAX test window: one-step-ahead metrics, plus the 20-day backtest (backtest_*_h20). MAPE/MPE are in %, the other metrics are in index points.

metric (test)	value
me	2.7991
rmse	72.9456
mae	57.7989
mpe	0.0456
mape	1.0469
mase	3.1969
backtest_mae_h20	174.2672
backtest_mape_h20	3.1223

Experiments

Selection rule (decided before, and independent of, the test window; test metrics are never used): rank the classical models (ARIMA, Holt-Winters, VAR) by one-step validation MAPE; among all models within 2 % (relative) of the best, deploy the one with the fewest fitted parameters (ties go to the lower validation MAPE). The classical models differ by hundredths of a percent, so a 164-parameter VAR(10) should only be deployed when no small model is close. Fitted parameters: Holt-Winters 4, ARIMA 6-7 (AR + MA coefficients + noise variance), VAR(10) 164. Result: VAR(10) has the lowest validation MAPE on every index, but Holt-Winters is within 2 % of it on DAX (0.967 vs 0.952, +1.6 %), SMI (+1.5 %) and FTSE (+0.4 %), so Holt-Winters is deployed there. On CAC, Holt-Winters is 6.3 % behind (0.898 vs 0.845) and ARIMA 8.4 %, so VAR(10) stays. Naive persistence has no fitted parameters and is within 2 % on DAX too, but it is the reference, not a candidate.

DAX, every variant. Validation MAPE (through the rule above) picks the deployed model; test columns are the R accuracy() table on the last 100 days (one-step-ahead). MASE is scaled by the in-sample naive MAE of the calm early 1990s, which is why it is > 1 for every model, naive included.

variant	val MAPE %	test ME	test RMSE	test MAE	test MPE %	test MAPE %	test MASE
ARIMA(2,1,3)	0.995	4.01	72.47	58.22	0.067	1.054	3.22
Holt-Winters (deployed)	0.967	2.80	72.95	57.80	0.046	1.047	3.20
VAR(10)	0.952	0.95	74.05	58.60	0.009	1.061	3.24
MLP	2.757	124.96	168.21	148.76	2.220	2.657	8.23
LSTM	1.919	68.98	171.69	147.43	1.169	2.622	8.15
Naive persistence	0.970	4.07	73.01	57.88	0.068	1.048	3.20

Every index, validation MAPE / test MAPE (%), deployed in bold. VAR has the lowest validation MAPE on all four indices, yet the lowest test MAPE on none: the differences between the classical models (and naive) are a few hundredths of a percent, well inside day-to-day noise. That Holt-Winters is also the best classical model on the DAX and FTSE test windows is not why it was chosen; the rule never reads the test columns.

index	ARIMA	Holt-Winters	VAR	MLP	LSTM	Naive persistence	deployed
DAX	0.995 / 1.054	0.967 / 1.047	0.952 / 1.061	2.757 / 2.657	1.919 / 2.622	0.970 / 1.048	Holt-Winters
SMI	0.844 / 0.934	0.771 / 0.943	0.760 / 0.951	3.280 / 2.954	1.533 / 2.246	0.778 / 0.943	Holt-Winters
CAC	0.916 / 0.989	0.898 / 0.969	0.845 / 1.012	1.737 / 1.596	1.212 / 1.282	0.903 / 0.968	VAR(10)
FTSE	0.808 / 0.795	0.779 / 0.788	0.776 / 0.812	2.906 / 2.265	1.413 / 2.475	0.782 / 0.788	Holt-Winters

Multi-step backtest (what the Space serves): MAE in index points of 20-business-day forecasts from 9 origins in the test window, deployed in bold. Over this short window, the deployed Holt-Winters beats naive on DAX (174.3 vs 195.7) and SMI (205.8 vs 208.4) and loses on FTSE (160.4 vs 148.1); the deployed VAR(10) loses to naive on CAC (132.7 vs 129.8). Holt-Winters is the best on DAX and SMI, ARIMA on CAC, and naive persistence on FTSE.

index	Holt-Winters MAE	ARIMA MAE	VAR MAE	Naive persistence MAE	origins
DAX	174.3	195.0	184.5	195.7	9 x 20 days
SMI	205.8	216.0	239.8	208.4	9 x 20 days
CAC	135.8	124.7	132.7	129.8	9 x 20 days
FTSE	160.4	149.5	166.4	148.1	9 x 20 days

The neural nets are much worse here (test MAPE 1.3-3.0 %) than in the original course run. Their scaler is fit on the fit window only, and the 1998 validation / test levels lie far above anything in it (DAX tops out at about 4,460 in the fit window; the test window reaches 6,186). The nets cannot extrapolate, and on DAX early stopping keeps the weights from epoch 6 (MLP) and epoch 2 (LSTM). The classical models work on levels and differences and do not have this problem.

Limitations

• Near random walk. One-step test errors of every classical model are within 0.05 MAPE points of "tomorrow = today" on every index (0.02 on DAX). This model does not predict the market; it is a teaching example of the classical workflow.
• Selection is noisy. With 100 validation days, the ranking of ARIMA / Holt-Winters / VAR can flip between windows. VAR won validation on every index but was the weakest classical model on the test window, which is why the rule prefers a 4-parameter model whenever its validation MAPE is within 2 %.
• Synthetic dates. Weekdays from 1991-07-01, no holidays. Treat the calendar dates as approximate.
• Old, closed data. 1991-1998 index closes only. Cutoffs after 1998-08-14 just forecast from the last row; there is no live data.
• No intervals. Point forecasts only. Uncertainty grows fast with the horizon (the 20-day backtest MAPE is about 3 % vs about 1 % one step ahead).
• ARIMA(2,1,3) on DAX did not converge in training (its AR and MA roots nearly cancel). It is kept because it had the lowest AIC, and its forecasts are almost identical to naive.