My project involves trying to predict the sales quantity for a specific item across a whole year. I've used the LightGBM package for making the predictions. The params I've set for it are as follows:

params = {
'nthread': 10,
'max_depth': 5, #DONE
'task': 'train',
'boosting_type': 'gbdt',
'objective': 'regression_l1',
'metric': 'mape', # this is abs(a-e)/max(1,a)
'num_leaves': 2, #DONE
'learning_rate': 0.2180, #DONE
'feature_fraction': 0.9, #DONE
'bagging_fraction': 0.990, #DONE
'bagging_freq': 1, #DONE
'lambda_l1': 3.097758978478437, #DONE
'lambda_l2': 2.9482537987198496, #DONE
'verbose': 1,
'min_child_weight': 6.996211413900573,
'min_split_gain': 0.037310344962162616,
'min_data_in_bin': 1, #DONE
'min_data_in_leaf':2, #DONE
'num_boost_round': 1, #DONE
'max_bin': 7, #DONE
'extra_trees': True, #DONE

My dataset consists of daily sales data (columns= date, quantity) for the years 2017, 2018, 2019 and 3 months of 2020. I've been trying to use the 2017 and 2018 data for training and cross-validation and trying to test it for 2019 data. However my predictions for the year is way off the mark while considering the quantities on a weekly, monthly, quarterly or yearly basis (error ~ 40-50%)(I've tuned the params to bring the error down to this values). Moreover while considering the predictions, my r2_score is giving me a negative value of around -2.9148426301633803. Any suggestions on what can be done to make it better?

Script for lightgbm:

lgb_train = lgb.Dataset(train_x, train_y)
lgb_valid = lgb.Dataset(test_x, test_y)
model = lgb.train(params, lgb_train, \
                  valid_sets=[lgb_train, lgb_valid],\
test_df_pred = df[(df.date >= '2019-01-01') & (df.date < '2020-01-01')]
#test_df_pred = df[(df.date >= '2019-01-01') & (df.date < '2019-02-01')]
#test_df_pred = df[(df.date >= '2019-01-15') & (df.date < '2019-01-22')]
test_df_pred['month'] = test_df_pred['date'].dt.month
test_df_pred['day'] = test_df_pred['date'].dt.dayofweek
test_df_pred['year'] = test_df_pred['date'].dt.year
col = [i for i in test_df_pred.columns if i not in ['date','id', 'qty']]
y_test_pred = model.predict(test_df_pred[col])
test_df_pred['qty_pred'] = y_test_pred
mse = mean_squared_error(y_true=test_df_pred['qty'], y_pred=test_df_pred['qty_pred'])
mae = mean_absolute_error(y_true=test_df_pred['qty'], y_pred=test_df_pred['qty_pred'])
mape = mean_absolute_percentage_error(y_true=test_df_pred['qty'], y_pred=test_df_pred['qty_pred'])
qty = test_df_pred.qty.sum()
qty_pred = test_df_pred.qty_pred.sum()
diff = qty_pred - qty

enter image description here enter image description here

enter image description here

  • 1
    $\begingroup$ It would be much more helpful if you provide the structure and content of your data and your script. $\endgroup$ Commented Nov 3, 2020 at 15:27
  • 1
    $\begingroup$ I've added the script and a screenshot of my data. Please let me know what you think. $\endgroup$ Commented Nov 3, 2020 at 18:43

1 Answer 1


I assume you are new to the field, thus, I would suggest using tutorials to achieve your goal. Because what you did is completely wrong and your approach is incorrect. I guess you want to model the sales as time series without using any predictor instead you want to model future values by looking at the past values. To achieve that, you need to use algorithms like ARIMA, exponential smoothing, etc. Here what you have done is trying to correlate the year, month, and day with the sales, which does not possess any information about the sale as expected (also you decoded it wrongly). Thus, your performance metric shows you a negative result. As a reference, check these which are similar to your problem. Source1, Source2, Source3. These will solve your issue.

  • $\begingroup$ Can you please explain what went wrong with the decoding as you mentioned? $\endgroup$ Commented Nov 4, 2020 at 7:58
  • $\begingroup$ I actually used LightGBM because I thought later on I could include additional features like holidays etc to help with prediction. Otherwise yes, A better option would've been to do for ARIMA or prophet $\endgroup$ Commented Nov 4, 2020 at 8:05
  • $\begingroup$ You should not include the date as your predictor because it does not have any correlation with your target value. If you do not have any predictor then use target itself, that is, target feature's past values to predict future. However, if you have features firstly analyze their bivariate relationship with the target value then add the ones that have relationship with the target value to the model. You can also, do vice versa, using all your variables in a model then drop the ones which indicates no relationship with target. But, I would suggest first approach. $\endgroup$ Commented Nov 4, 2020 at 12:47
  • $\begingroup$ You decoded the day starting from 0, also your first day is 2 (2017-01-02). Again, I don't think the date can somehow be your predictor, but still I am saying your both shifted your decoding also started from 0. $\endgroup$ Commented Nov 4, 2020 at 12:50
  • $\begingroup$ so what u suggest would be good features to help in prediction which can derived from sales quantity? Like i can add features for holidays and other data but any other suggestions? $\endgroup$ Commented Nov 9, 2020 at 11:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.